RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.
Go agent-harnessagent-native-databaseagentic-aiagentic-retrievalagentic-searchaiai-agentscontext-enginecontext-engineeringcontext-managementharness-engineeringknowledge-compilationragretrieval-augmented-generationretrieval-harnesssearch-harness
refactor(deps): use Go dependency downloader as sole entrypoint (#20360)
jay777214071451ea4b9
diff --git a/.dockerignore b/.dockerignoreindex 1b4707c99e..941513916c 100644--- a/.dockerignore+++ b/.dockerignore@@ -35,14 +35,12 @@ target/# ── ragflow_deps build context (built as a separate image, mounted ──# ── from infiniflow/ragflow_deps:latest by the main Dockerfile) ──# Excluding the directory contents keeps the main build context small-# regardless of which deps files download_deps.py currently fetches.+# regardless of which resources are prepared for the dependency image.# The deps image is built from this directory with:# cd ragflow_deps && docker build -f Dockerfile -t infiniflow/ragflow_deps .-# The two ORT version-pin files are re-included: build.sh's-# check_ort_version_consistency (run by Dockerfile_go's `--go` build) greps-# them. Re-including requires NOT excluding the directory itself.+# The Go downloader is re-included because build.sh reads its ORT version pin.+# Re-including requires NOT excluding the directory itself.ragflow_deps/*-!ragflow_deps/download_go_deps.py!ragflow_deps/download_deps.py# ── IDE and editor config ──────────────────────────────────────────────────diff --git a/.github/workflows/sep-tests.yml b/.github/workflows/sep-tests.ymlindex 5645114c5a..dd168e6033 100644--- a/.github/workflows/sep-tests.yml+++ b/.github/workflows/sep-tests.yml@@ -923,7 +923,7 @@ jobs:run: |set -euo pipefailuv sync --python 3.13 --group test --frozen- uv run pytest ragflow_deps/test_download_go_deps.py ragflow_deps/test_tokenizer_assets.py -q --durations=100+ uv run pytest ragflow_deps/test_download_deps.py ragflow_deps/test_tokenizer_assets.py -q --durations=100- name: Build ragflow:nightlyrun: |diff --git a/AGENTS.md b/AGENTS.mdindex a0016b30af..3911b37149 100644--- a/AGENTS.md+++ b/AGENTS.md@@ -53,7 +53,7 @@ Use this file as the local operating guide for the current codebase. Prefer the- `web/`: frontend application.- `docker/`: local and production compose files.- `sdk/python/`: legacy Python client SDK pending deletion.-- `ragflow_deps/download_go_deps.py`: current Python helper for Go native libraries and model files; replace it before deleting it. `build.sh` currently reads its ONNX Runtime version pin.+- `ragflow_deps/download_deps.py`: Python helper for Go native libraries, DeepDoc models, and tokenizer assets. Replace it with non-Python preparation before removing Python dependencies.- `test/`: contains Python tests, including tests that import the legacy Python implementation directly. Migrate needed coverage to Go tests or frontend tests, then remove the Python tests; keep only tests for supported code.## Go-Specific Rules@@ -129,7 +129,7 @@ bash build.sh --test-manual # local opt-in onlybash build.sh --test-all # integration + e2e; excludes manual```-Native libraries and Go DeepDoc models currently come from `ragflow_deps/download_go_deps.py`; see `README_zh.md` for its isolated setup. The Go tokenizer also needs `ragflow_deps/cl100k_base.tiktoken`, which the current Python dependency pipeline provides. These are build/runtime resource preparation steps, not a Python server requirement.+Native libraries, Go DeepDoc models, tokenizer assets, and `ragflow_deps/cl100k_base.tiktoken` come from `ragflow_deps/download_deps.py`; see `README_zh.md` for its isolated setup. The downloader installs stagehand into the SDK cache using the version in `go.mod`. Build the Go resource image with `cd ragflow_deps && docker build -f Dockerfile -t infiniflow/ragflow_deps:latest .` after downloading the resources.### Frontend```bashdiff --git a/Dockerfile b/Dockerfileindex c3e7633f25..55518459ab 100644--- a/Dockerfile+++ b/Dockerfile@@ -30,7 +30,6 @@ RUN --mount=type=bind,from=infiniflow/ragflow_deps:latest,source=/huggingface.co--exclude='layout.manual.ort' \--exclude='layout.paper.ort' \-cf - \- /huggingface.co/InfiniFlow/text_concat_xgb_v1.0 \/huggingface.co/InfiniFlow/deepdoc \| tar -xf - --strip-components=3 -C /ragflow/internal/rag/res/deepdoc && \#ln -s layout.onnx /ragflow/internal/rag/res/deepdoc/layout.laws.onnx && \@@ -186,9 +185,9 @@ COPY internal internalCOPY cmd cmdCOPY build.sh ./# build.sh's check_ort_version_consistency (run via `./build.sh --go`) greps the-# ORT version pins from these files; without them the --go build fails with+# ORT version pin from this file; without it the --go build fails with# "could not parse the ONNX Runtime version from one of the pinned locations".-COPY ragflow_deps/download_go_deps.py ragflow_deps/download_deps.py ./ragflow_deps/+COPY ragflow_deps/download_deps.py ./ragflow_deps/COPY Dockerfile ./# ONNX Runtime static archives: build.sh's _seed_from_system looks for the ORTdiff --git a/README.md b/README.mdindex b53c059436..694cd3465f 100644--- a/README.md+++ b/README.md@@ -255,12 +255,12 @@ Install the Go version specified in `go.mod` (currently Go 1.27), Clang 20, LLDcd ragflow```-2. Prepare native libraries and model files with the [Go dependency download script](./ragflow_deps/download_go_deps.py), then build the Go services:+2. Prepare native libraries, model files, and tokenizer assets with the [Go dependency download script](./ragflow_deps/download_deps.py), then build the Go services:```bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub- /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+ /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.pybash build.sh --all```diff --git a/README_ar.md b/README_ar.mdindex cfa7b18e6c..4155e49b1f 100644--- a/README_ar.md+++ b/README_ar.md@@ -281,7 +281,7 @@```bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub- /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+ /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.pybash build.sh --all```diff --git a/README_fr.md b/README_fr.mdindex d9707706be..f8f6b2fae1 100644--- a/README_fr.md+++ b/README_fr.md@@ -269,7 +269,7 @@ Pour changer de moteur de documents, modifier la configuration, redémarrer les```bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub- /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+ /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.pybash build.sh --all```diff --git a/README_id.md b/README_id.mdindex 8baf7f0442..de3d5c93e5 100644--- a/README_id.md+++ b/README_id.md@@ -277,7 +277,7 @@ Untuk mengganti mesin dokumen, mengubah konfigurasi, memulai ulang layanan, sert```bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub- /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+ /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.pybash build.sh --all```diff --git a/README_ja.md b/README_ja.mdindex bb2d434ede..2e8b71808e 100644--- a/README_ja.md+++ b/README_ja.md@@ -259,7 +259,7 @@ Go版のDockerデプロイでは `docker/.env` と `docker/docker-compose.yml````bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub- /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+ /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.pybash build.sh --all```diff --git a/README_ko.md b/README_ko.mdindex 68b7044c03..122403ba4c 100644--- a/README_ko.md+++ b/README_ko.md@@ -259,7 +259,7 @@ Go Docker 배포는 `docker/.env`와 `docker/docker-compose.yml`을 사용하며```bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub- /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+ /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.pybash build.sh --all```diff --git a/README_pt_br.md b/README_pt_br.mdindex 49bec0c864..ad017d5ae3 100644--- a/README_pt_br.md+++ b/README_pt_br.md@@ -265,7 +265,7 @@ Para trocar o mecanismo de documentos, alterar configurações, reiniciar servi```bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub- /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+ /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.pybash build.sh --all```diff --git a/README_ru.md b/README_ru.mdindex d09cc90d22..fd3b7371e3 100644--- a/README_ru.md+++ b/README_ru.md@@ -275,7 +275,7 @@```bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub- /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+ /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.pybash build.sh --all```diff --git a/README_tr.md b/README_tr.mdindex 0395c24b43..bbb20b8fd4 100644--- a/README_tr.md+++ b/README_tr.md@@ -277,7 +277,7 @@ Belge motorunu değiştirme, yapılandırma değişikliklerinden sonra hizmetler```bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub- /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+ /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.pybash build.sh --all```diff --git a/README_tzh.md b/README_tzh.mdindex ade350faf6..b6e2ff903b 100644--- a/README_tzh.md+++ b/README_tzh.md@@ -272,12 +272,12 @@ Go 版 Docker 部署使用 `docker/.env` 和 `docker/docker-compose.yml`,以 Kgit clone https://github.com/infiniflow/ragflow.gitcd ragflow```-2. 使用 [Go 依賴下載腳本](./ragflow_deps/download_go_deps.py)準備原生程式庫和模型檔案,再編譯 Go 服務:+2. 使用 [Go 依賴下載腳本](./ragflow_deps/download_deps.py)準備原生程式庫、模型檔案和 tokenizer 資源,再編譯 Go 服務:```bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub- /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+ /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.pybash build.sh --all```diff --git a/README_zh.md b/README_zh.mdindex 65a18287da..abc0bf0ba9 100644--- a/README_zh.md+++ b/README_zh.md@@ -279,12 +279,12 @@ Go 版 Docker 部署使用 `docker/.env` 和 `docker/docker-compose.yml`,使cd ragflow```-2. 按 [Go 依赖下载脚本](./ragflow_deps/download_go_deps.py)准备原生库和模型文件,再编译 Go 服务:+2. 按 [Go 依赖下载脚本](./ragflow_deps/download_deps.py)准备原生库、模型文件和 tokenizer 资源,再编译 Go 服务:```bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub- /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+ /tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.pybash build.sh --all```diff --git a/build.sh b/build.shindex e3a89964a4..1dd09bc8cd 100755--- a/build.sh+++ b/build.sh@@ -42,8 +42,7 @@ PDF_OXIDE_VERSION="0.3.73"# (--undefined=OrtGetApiBase + --dynamic-list, no --whole-archive); OrtGetApiBase# is then resolved via dlopen(NULL) (the process-global symbol table, not the# executable's own path), so no libonnxruntime.so is needed at runtime.-# Downloaded by ragflow_deps/download_go_deps.py (and-# ragflow_deps/download_deps.py) into onnxruntime/static_lib.+# Downloaded by ragflow_deps/download_deps.py into onnxruntime/static_lib.ONNXRUNTIME_STATIC_PREFIX="${HOME}/ragflow-native-libs/onnxruntime/static_lib"# Copy a dependency from the system pre-seed directory to the user cache.@@ -219,26 +218,24 @@ check_go_deps() {# The in-process (Go) DeepDoc backend statically links libonnxruntime.a built# from ONE exact ORT release; a mismatch links a wrong/missing .a and only fails# at runtime (dlopen(NULL) can't find OrtGetApiBase). The version is pinned in-# four Go-side locations that must all agree.+# three Go-side locations that must all agree.check_ort_version_consistency() {print_section "Checking ONNX Runtime version consistency"- local env_go d1 d2 dockerfile+ local env_go deps dockerfileenv_go="$(grep -m1 -E 'DeepDocORTVersion[[:space:]]*=[[:space:]]*"' "${PROJECT_ROOT}/internal/common/environments.go" | sed -E 's/.*"([^"]+)".*/\1/')"- d1="$(grep -m1 -E '^ORT_VERSION[[:space:]]*=[[:space:]]*"' "${PROJECT_ROOT}/ragflow_deps/download_go_deps.py" | sed -E 's/.*"([^"]+)".*/\1/')"- d2="$(grep -m1 -E '^ORT_VERSION[[:space:]]*=[[:space:]]*"' "${PROJECT_ROOT}/ragflow_deps/download_deps.py" | sed -E 's/.*"([^"]+)".*/\1/')"+ deps="$(grep -m1 -E '^ORT_VERSION[[:space:]]*=[[:space:]]*"' "${PROJECT_ROOT}/ragflow_deps/download_deps.py" | sed -E 's/.*"([^"]+)".*/\1/')"dockerfile="$(grep -m1 -E 'ARG[[:space:]]+ORT_VERSION=' "${PROJECT_ROOT}/Dockerfile" | sed -E 's/.*ORT_VERSION=([0-9][^"[:space:]]*).*/\1/')"- if [ -z "$env_go" ] || [ -z "$d1" ] || [ -z "$d2" ] || [ -z "$dockerfile" ]; then+ if [ -z "$env_go" ] || [ -z "$deps" ] || [ -z "$dockerfile" ]; thenecho -e "${RED}Error: could not parse the ONNX Runtime version from one of the pinned locations${NC}" >&2exit 1fi- if [ "$env_go" != "$d1" ] || [ "$env_go" != "$d2" ] || [ "$env_go" != "$dockerfile" ]; then+ if [ "$env_go" != "$deps" ] || [ "$env_go" != "$dockerfile" ]; thenecho -e "${RED}Error: ONNX Runtime version is inconsistent — fix before building:${NC}" >&2printf ' %-10s %s\n' "$env_go" "internal/common/environments.go:DeepDocORTVersion"- printf ' %-10s %s\n' "$d1" "ragflow_deps/download_go_deps.py:ORT_VERSION"- printf ' %-10s %s\n' "$d2" "ragflow_deps/download_deps.py:ORT_VERSION"+ printf ' %-10s %s\n' "$deps" "ragflow_deps/download_deps.py:ORT_VERSION"printf ' %-10s %s\n' "$dockerfile" "dockerfile:ARG ORT_VERSION"exit 1fi@@ -302,7 +299,7 @@ check_onnxruntime_deps() {echo -e "${RED}Error: no ONNX Runtime ${ort_version} static lib under $ONNXRUNTIME_STATIC_PREFIX${NC}" >&2echo " available: ${avail:-<none>}" >&2echo " DeepDocORTVersion=${ort_version}; fetch it with:" >&2- echo " uv run python3 ragflow_deps/download_go_deps.py" >&2+ echo " uv run python3 ragflow_deps/download_deps.py" >&2echo " or pre-seed /opt/ragflow-native-libs/onnxruntime (CI image)." >&2return 1fi@@ -316,8 +313,8 @@ check_onnxruntime_deps() {# or GitHub was unreachable (curl had no connect timeout).## A mismatch means the local copy is corrupt or stale; we only WARN (not- # fail) so a one-off re-issue cannot break the build. The download scripts- # (ragflow_deps/download_go_deps.py, download_deps.py) still verify against+ # fail) so a one-off re-issue cannot break the build. The downloader+ # ragflow_deps/download_deps.py still verifies against# the published sidecar at download time, so a re-issued archive is picked# up on the next download.local asset="onnxruntime-v${ort_version}-linux-x86_64.zip"@@ -332,7 +329,7 @@ check_onnxruntime_deps() {echo " local sha256: $actual" >&2echo " The local copy may be corrupt or stale. Refresh it with:" >&2echo " rm -f ${zip_path}" >&2- echo " uv run python3 ragflow_deps/download_go_deps.py # or ragflow_deps/download_deps.py" >&2+ echo " uv run python3 ragflow_deps/download_deps.py" >&2elseecho " ✓ ${asset} sha256 matches pinned digest"fi@@ -357,7 +354,7 @@ check_office_oxide_deps() {if [ ! -f "$lib_path" ] || [ ! -f "$header_path" ]; thenecho -e "${RED}Error: office_oxide native library not found${NC}"echo " Expected: ${lib_path}"- echo " Run: uv run python3 ragflow_deps/download_go_deps.py"+ echo " Run: uv run python3 ragflow_deps/download_deps.py"echo " Or manually download: https://github.com/yfedoseev/office_oxide/releases/download/v${OFFICE_OXIDE_VERSION}/native-linux-x86_64.tar.gz"exit 1fi@@ -371,12 +368,12 @@ check_office_oxide_deps() {found_version=$(strings "$lib_path" 2>/dev/null | grep -E "^0\.[0-9]+\.[0-9]+$" | head -1)echo -e "${YELLOW}office_oxide native lib version mismatch (required v${OFFICE_OXIDE_VERSION}; found: ${found_version:-unknown}); refreshing from download${NC}"rm -rf "${OFFICE_OXIDE_PREFIX}" "${PROJECT_ROOT}/office_oxide-linux-x86_64.tar.gz"- if ! (cd "${PROJECT_ROOT}" && uv run python3 ragflow_deps/download_go_deps.py); then+ if ! (cd "${PROJECT_ROOT}" && uv run python3 ragflow_deps/download_deps.py); thenecho -e "${RED}Error: office_oxide native lib version mismatch${NC}"echo " Required: v${OFFICE_OXIDE_VERSION}; found: ${found_version:-unknown}"echo " A stale lib silently loses PPT97 (.ppt) slide content. Refresh:"echo " rm -rf ~/ragflow-native-libs/office_oxide ragflow_deps/office_oxide-linux-x86_64.tar.gz"- echo " uv run python3 ragflow_deps/download_go_deps.py"+ echo " uv run python3 ragflow_deps/download_deps.py"exit 1filib_path="${OFFICE_OXIDE_PREFIX}/lib/${lib_file}"@@ -409,7 +406,7 @@ check_pdfium_deps() {echo " pdfium (static) not found"echo " Expected: ${lib_path}"- echo " Run: uv run python3 ragflow_deps/download_go_deps.py"+ echo " Run: uv run python3 ragflow_deps/download_deps.py"echo " Or: curl -fsSL https://github.com/kognitos/pdfium-static/releases/download/chromium%2F${PDFIUM_STATIC_VERSION}/pdfium-linux-x64-static.tgz | tar xz -C ${PDFIUM_STATIC_PREFIX}"return 1}@@ -515,7 +512,7 @@ check_pdf_oxide_deps() {echo " Required: v${PDF_OXIDE_VERSION}; found: ${found_version:-unknown}"echo " A stale lib silently reverts PDF parsing fixes. Refresh:"echo " rm -rf ${PDF_OXIDE_PREFIX} ragflow_deps/pdf_oxide-go-ffi-linux-amd64.tar.gz"- echo " uv run python3 ragflow_deps/download_go_deps.py"+ echo " uv run python3 ragflow_deps/download_deps.py"return 1;;esac@@ -523,7 +520,7 @@ check_pdf_oxide_deps() {echo " pdf_oxide (static) not found"echo " Expected: ${lib_path}"- echo " Run: uv run python3 ragflow_deps/download_go_deps.py"+ echo " Run: uv run python3 ragflow_deps/download_deps.py"echo " Or: curl -fsSL https://github.com/yfedoseev/pdf_oxide/releases/download/v${PDF_OXIDE_VERSION}/pdf_oxide-go-ffi-linux-amd64.tar.gz | tar xz -C ${PDF_OXIDE_PREFIX}"return 1}@@ -653,12 +650,12 @@ build_go() {echo " -Wl,--dynamic-list). Without it the binary compiles but dies" >&2echo " at startup with a fatal 'no in-process DeepDoc backend serving'." >&2echo " Fetch the static libs with:" >&2- echo " uv run python3 ragflow_deps/download_go_deps.py" >&2+ echo " uv run python3 ragflow_deps/download_deps.py" >&2echo " or pre-seed them at /opt/ragflow-native-libs/onnxruntime (CI image)." >&2echo " This production binary must statically include ORT — there is no" >&2echo " ORT-free build path. ORT ends up unlinked only via one of:" >&2echo " - the ORT static_lib dir was never seeded: run" >&2- echo " 'uv run python3 ragflow_deps/download_go_deps.py', or pre-seed" >&2+ echo " 'uv run python3 ragflow_deps/download_deps.py', or pre-seed" >&2echo " /opt/ragflow-native-libs/onnxruntime as the CI image does;" >&2echo " - setup_cgo_env did not add libonnxruntime to CGO_LDFLAGS." >&2return 1@@ -1096,7 +1093,7 @@ DEPENDENCIES:- cmake >= 4.0- go >= 1.26.4- clang++ with C++20 support- - office_oxide native library (download with: uv run python3 ragflow_deps/download_go_deps.py)+ - office_oxide native library (download with: uv run python3 ragflow_deps/download_deps.py)- lld (Linux only): sudo apt install lld-20 && sudo ln -s /usr/bin/ld.lld-20 /usr/bin/ld.lld- pcre2 development files- Debian/Ubuntu: libpcre2-dev@@ -1198,7 +1195,7 @@ main() {fi# The in-process (Go) DeepDoc backend needs the .ort weights. Default# MODEL_DIR to the canonical repo model dir (internal/rag/res/deepdoc) so a- # local run needs no MODEL_DIR export after `download_go_deps.py`.+ # local run needs no MODEL_DIR export after `download_deps.py`.REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"export MODEL_DIR="${MODEL_DIR:-$REPO_ROOT/internal/rag/res/deepdoc}"run_go_tests_tagged "cgo integration" "${pkgs[@]}"diff --git a/cmd/ragflow_server.go b/cmd/ragflow_server.goindex 2b66a6d5d8..305268d7f6 100644--- a/cmd/ragflow_server.go+++ b/cmd/ragflow_server.go@@ -1674,10 +1674,14 @@ func logTokenizerCounters() {if len(unavailable) > 0 {common.Warn("embedding tokenizers unavailable; models that declare them cannot be ingested until the asset is restored",zap.Strings("unavailable", unavailable),- zap.String("hint", "run `uv run ragflow_deps/download_go_deps.py`, or set "+common.EnvModelAssetsDir+" to a directory holding them"))+ zap.String("hint", "run `uv run ragflow_deps/download_deps.py`, or set "+common.EnvModelAssetsDir+" to a directory holding them"))}}+// resolveDeepDocModelDir picks the model directory: the explicit DEEPDOC_MODEL_DIR+// env, else the RAGFlow default model dir (internal/rag/res/deepdoc),+// else the assets fetched by ragflow_deps/download_deps.py. The first+// candidate that actually contains the required weights wins.func resolveDeepDocModelDir() string {if v := strings.TrimSpace(common.GetEnv(common.EnvDeepDocModelDir)); v != "" {return vdiff --git a/docker/launch_backend_service.sh b/docker/launch_backend_service.shindex 8fa2de7038..eba2d5f96e 100755--- a/docker/launch_backend_service.sh+++ b/docker/launch_backend_service.sh@@ -72,9 +72,7 @@ STOP=falsePIDS=()# Set the path to the NLTK data directory-# download_deps.py downloads NLTK data into ragflow_deps/nltk_data (see-# ragflow_deps/download_deps.py); point NLTK_DATA there directly instead of a-# stale top-level ./nltk_data so this matches what's actually populated.+# Use the repository's NLTK data directory when it is provisioned.export NLTK_DATA="$(pwd)/ragflow_deps/nltk_data"# Function to handle termination signalsdiff --git a/docs/develop/launch_ragflow_from_source.md b/docs/develop/launch_ragflow_from_source.mdindex d22e488772..dc1752c09a 100644--- a/docs/develop/launch_ragflow_from_source.md+++ b/docs/develop/launch_ragflow_from_source.md@@ -27,7 +27,7 @@ All long-running RAGFlow processes started below use `bin/ragflow_server`.- Go 1.27 or later, as declared in `go.mod` (check `go version`).- CMake 4.0 or later, Clang 20, LLD 20, and PCRE2 development headers.- Node.js 18.20.4 or later and npm for the frontend.-- Python 3.10 or later only for `ragflow_deps/download_go_deps.py`, which downloads the native libraries and model resources required by the Go build.+- Python 3.10 or later only for `ragflow_deps/download_deps.py`, which downloads the native libraries and model resources required by the Go build.See the [Docker installation guide](https://docs.docker.com/engine/install/) if Docker is not installed. Use the Go version declared in `go.mod` and the compiler versions listed above when preparing the build environment.@@ -59,10 +59,12 @@ Download the native libraries and Go DeepDoc model weights with a small, isolate```bashpython3 -m venv /tmp/ragflow-go-download-venv/tmp/ragflow-go-download-venv/bin/python -m pip install requests huggingface-hub-/tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_go_deps.py+/tmp/ragflow-go-download-venv/bin/python ragflow_deps/download_deps.py```-The downloader fetches the static libraries needed by `build.sh`, plus `det.ort`, `layout.ort`, `tsr.ort`, `rec.ort`, and `ocr.res` into `internal/rag/res/deepdoc/`. These files are required by the in-process Go DeepDoc backend. Keep the server's working directory at the repository root so it can find them automatically; if you launch it elsewhere, set `DEEPDOC_MODEL_DIR` to the absolute path of `rag/res/deepdoc`.+The downloader fetches the static libraries needed by `build.sh`, plus `det.ort`, `layout.ort`, `tsr.ort`, `rec.ort`, and `ocr.res` into `internal/rag/res/deepdoc/`. These files are required by the in-process Go DeepDoc backend. Keep the server's working directory at the repository root so it can find them automatically; if you launch it elsewhere, set `DEEPDOC_MODEL_DIR` to the absolute path of `internal/rag/res/deepdoc`.++It also downloads `ragflow_deps/cl100k_base.tiktoken` and the embedding tokenizer assets, and installs the stagehand driver into `${XDG_CACHE_HOME:-$HOME/.cache}/stagehand/lib/go_<SDK version>/`. Native archives support Linux x86_64.The isolated download environment can be removed after the resources have been prepared.@@ -72,7 +74,7 @@ Build the C++ bindings and Go binaries:bash build.sh --all```-For a smaller production binary, use `bash build.sh --strip --all` instead. The build script also tries to place `ragflow_deps/cl100k_base.tiktoken` in the repository. If it reports that the BPE table could not be provisioned, download it before starting the server:+For a smaller production binary, use `bash build.sh --strip --all` instead. The dependency downloader prepares `ragflow_deps/cl100k_base.tiktoken`. If the table is missing, download it before starting the server:```bashcurl -fsSL -o ragflow_deps/cl100k_base.tiktoken https://openaipublic.blob.core.windows.net/encodings/cl100k_base.tiktoken@@ -88,6 +90,16 @@ It should print API usage information and may exit with status 1. This command oThe Go/C++ build requires LLD 20. Confirm that `ld.lld --version` reports LLD 20 before building; merely installing `lld-20` is insufficient when an older `ld.lld` remains the system default. If the binary builds successfully but exits or crashes before reaching Go `main`, verify the selected linker and rebuild with LLD 20.+To prepare the resource image used by the Go Dockerfile, run the downloader above, then build from `ragflow_deps/`:++```bash+cd ragflow_deps+docker build -f Dockerfile -t infiniflow/ragflow_deps:latest .+cd ..+```++The downloader prepares the five Go DeepDoc files under `ragflow_deps/huggingface.co/InfiniFlow/deepdoc/` for the resource image, as well as `internal/rag/res/deepdoc/` for local use. The Go Dockerfile copies these models into the runtime model directory. Analyzer dictionaries and fonts are installed separately by the Go Dockerfile.+## 2. Start supporting servicesWith the default `conf/service_conf.yaml`, the host-run Go processes need Elasticsearch, MySQL, MinIO, NATS, Kvrocks, and ClickHouse. Start these services explicitly from the repository root:diff --git a/internal/agent/component/io/pdf_writer.go b/internal/agent/component/io/pdf_writer.goindex 9505f4e649..f5dc750fca 100644--- a/internal/agent/component/io/pdf_writer.go+++ b/internal/agent/component/io/pdf_writer.go@@ -59,7 +59,7 @@ var defaultPDFLatinFontPaths = []string{}var defaultPDFCJKFontPaths = []string{- // Use the language-specific Noto CJK Variable TTF shipped by Dockerfile_go.+ // Use the language-specific Noto CJK Variable TTF shipped in the runtime image.// gopdf uses its default variation instance; it does not expose variable// font axes, so this is used as the regular CJK face."/usr/local/share/fonts/truetype/noto/NotoSansCJKsc-VF.ttf",diff --git a/internal/common/environments.go b/internal/common/environments.goindex 91b6583fe0..f3ca53cc40 100644--- a/internal/common/environments.go+++ b/internal/common/environments.go@@ -292,11 +292,9 @@ const (// do set-membership checks); keep it stable so logs and diffs stay readable.//// External consumers that re-list these names must stay in sync:-// - ragflow_deps/download_go_deps.py re-lists them as DEEPDOC_MODEL_FILES+// - ragflow_deps/download_deps.py re-lists them as DEEPDOC_MODEL_FILES// (it fetches the files one by one, so it MUST be edited by hand when this// slice changes);-// - ragflow_deps/download_deps.py snapshots the whole InfiniFlow/deepdoc repo-// (so .ort lands in the model dir automatically — no FILES edit needed).var DeepDocModelFiles = []string{"det.ort","layout.ort",@@ -320,9 +318,9 @@ func HasModelFiles(dir string) bool {// DeepDocORTVersion is the onnxruntime native release the in-process (Go)// DeepDoc backend is built and tested against (e.g. "1.29.0"). It is ONE OF-// FOUR raw version declarations that must stay equal (the other three are-// ORT_VERSION in ragflow_deps/download_go_deps.py and ragflow_deps/download_deps.py,-// and ARG ORT_VERSION in Dockerfile_go) — NOT a single source of truth. The+// THREE raw version declarations that must stay equal (the other two are+// ORT_VERSION in ragflow_deps/download_deps.py and ARG ORT_VERSION in+// Dockerfile) — NOT a single source of truth. The// download URL and extracted dir name are built from those ORT_VERSION// constants, not from this one. The Go binding// (github.com/infiniflow/onnxruntime_go, the org mirror of yalue/onnxruntime_go)@@ -332,12 +330,11 @@ func HasModelFiles(dir string) bool {// the same minor line. ONNX Runtime is linked statically (libonnxruntime.a),// so there is no .so / SONAME at runtime.//-// To bump ORT, the four Go-side native pins above must change together+// To bump ORT, the three Go-side native pins above must change together// (drift breaks the static link or the runtime OrtGetApiBase lookup):// - DeepDocORTVersion (here, Go)-// - ORT_VERSION in ragflow_deps/download_go_deps.py// - ORT_VERSION in ragflow_deps/download_deps.py-// - ARG ORT_VERSION in Dockerfile_go+// - ARG ORT_VERSION in Dockerfile//// Separately, keep these on the same ORT minor line but version them// independently of the Go native lib (see pyproject.toml / development.md):diff --git a/internal/deepdoc/native/run.sh b/internal/deepdoc/native/run.shindex 8dd6d7b06e..d8bcffc808 100755--- a/internal/deepdoc/native/run.sh+++ b/internal/deepdoc/native/run.sh@@ -15,7 +15,7 @@ cd "$SCRIPT_DIR"# dlopen(NULL); no ORT_LIB is needed.## Default MODEL_DIR to the canonical repo model dir (internal/rag/res/deepdoc) so running-# `download_go_deps.py` is enough — no MODEL_DIR export is required for local+# `download_deps.py` is enough — no MODEL_DIR export is required for local# Go DeepDoc runs. Override with `MODEL_DIR=... bash run.sh` if needed.REPO_ROOT="$(cd "$SCRIPT_DIR/../../.." && pwd)"MODEL_DIR="${MODEL_DIR:-$REPO_ROOT/internal/rag/res/deepdoc}"diff --git a/internal/development.md b/internal/development.mdindex f40b1b7d37..3bec130058 100644--- a/internal/development.md+++ b/internal/development.md@@ -39,15 +39,28 @@ go version### 1.4 Install dependent library```shellsudo apt install libpcre2-dev-python3 ragflow_deps/download_go_deps.py+uv run ragflow_deps/download_deps.py```+Install `uv` before running the downloader; it installs the script's declared Python dependencies.++This downloader provisions Go native libraries, `.ort` weights, model tokenizer assets, and `cl100k_base.tiktoken`. It installs the local stagehand driver into the SDK's cache, using the SDK version from `go.mod`. Native archives support Linux x86_64.++Build the Go resource image from `ragflow_deps/` after preparing these files:++```shell+cd ragflow_deps+docker build -f Dockerfile -t infiniflow/ragflow_deps:latest .+cd ..+```++The image contains Go DeepDoc weights, tokenizer assets, the BPE table, and stagehand binaries. DeepDoc weights are prepared under `ragflow_deps/huggingface.co/InfiniFlow/deepdoc/` for packaging and under `internal/rag/res/deepdoc/` for local use.> **Note**: If you use IDEs like GoLand to run/debug directly (via Run/Debug buttons), or run `go build` / `go run` from command line, set these CGO environment variables:>> ```bash-> RAGFLOW_DEPS="${HOME}/ragflow-native-libs" # created by download_go_deps.py + download_deps.py+> RAGFLOW_DEPS="${HOME}/ragflow-native-libs" # created by download_deps.py> PLATFORM="linux_amd64" # or darwin_amd64, linux_arm64, darwin_arm64-> # NOTE: the ONNX Runtime static lib fetched by download_go_deps.py is+> # NOTE: the ONNX Runtime static lib fetched by download_deps.py is> # linux-x64 ONLY (onnxruntime-linux-x64-static_lib-*). On darwin_* / non-amd64> # PLATFORM values the production DeepDoc backend cannot be linked, so those> # PLATFORM examples cover office_oxide/pdfium/pdf_oxide only — ORT is a@@ -97,7 +110,7 @@ python3 ragflow_deps/download_go_deps.py> Since `build.sh` (`build_go`) now **fails fast** when ORT is absent from> `CGO_LDFLAGS`, this breakage surfaces at build time instead of at runtime. If> you see `Error: ONNX Runtime static libraries are not linked`, run-> `uv run python3 ragflow_deps/download_go_deps.py` (or pre-seed+> `uv run ragflow_deps/download_deps.py` (or pre-seed> `/opt/ragflow-native-libs/onnxruntime` as the CI runner image does). There is> no ORT-free production build path — if ORT is absent the binary fails at> startup, so the remedy is always to seed the static lib above, never to build@@ -107,8 +120,8 @@ python3 ragflow_deps/download_go_deps.py> that must stay in sync. Bumping it in one spot and not the others fails the> build with `Error: ONNX Runtime version is inconsistent`:> - `internal/common/environments.go` — `DeepDocORTVersion`-> - `Dockerfile_go` — `ARG ORT_VERSION`-> - `ragflow_deps/download_go_deps.py` and `ragflow_deps/download_deps.py` — `ORT_VERSION`+> - `Dockerfile` — `ARG ORT_VERSION`+> - `ragflow_deps/download_deps.py` — `ORT_VERSION`>> `build.sh` runs this consistency check automatically before the Go build> (through `check_go_deps`) and fails fast on any mismatch. Run it on demand@@ -120,13 +133,12 @@ python3 ragflow_deps/download_go_deps.py> **Note**: `build.sh` also guards ONNX Runtime **archive integrity** and the> **link cache**, because a silently-stale `.a` is easy to miss:> - `check_onnxruntime_deps` compares the local release ZIP under `ragflow_deps/`-> (`onnxruntime-v<ver>-linux-x86_64.zip`) against the published-> `onnxruntime-v<ver>-linux-x86_64.zip.sha256` sidecar — the same source of-> truth the download scripts use. If they differ (the archive was re-issued-> under the same tag/asset name, or an older download is present) the build-> fails fast with `Error: ONNX Runtime archive ... is stale`, printing the-> expected/actual sha256 and the exact refresh command-> (`rm -f <zip>` + `uv run python3 ragflow_deps/download_go_deps.py`). CI seeds+> (`onnxruntime-v<ver>-linux-x86_64.zip`) against the SHA-256 digest pinned in+> `build.sh`, without a network request. The downloader separately checks the+> published `onnxruntime-v<ver>-linux-x86_64.zip.sha256` sidecar. A local ZIP+> mismatch only emits a warning with the expected/actual SHA-256 and the+> refresh command (`rm -f <zip>` + `uv run ragflow_deps/download_deps.py`);+> the build continues checking and linking the extracted static libraries. CI seeds> ORT from `/opt` and has no local zip, so the check is skipped there (the bake> is authoritative). This matters because the `onnxruntime_go` binding reaches> ORT only through the OrtApi function-pointer table@@ -139,7 +151,7 @@ python3 ragflow_deps/download_go_deps.py> would otherwise silently reuse a stale linked binary. Stamping the path with> the archive's sha256 changes the flag string whenever the content changes →> automatic relink. If you ever see a stale-`.a` crash after an ORT re-issue,-> re-run `uv run python3 ragflow_deps/download_go_deps.py` (or `download_deps.py`)+> re-run `uv run ragflow_deps/download_deps.py`> so the stamp moves; a plain `go clean -cache` also forces it.### 1.5 Build RAGFlow@@ -174,9 +186,10 @@ supersedes the other, so do not delete one to "clean up".| Format | `.ort` | `.onnx` || Files | `det.ort`, `layout.ort`, `tsr.ort`, `rec.ort`, `ocr.res` | `det.onnx`, `layout.onnx`, `tsr.onnx`, `rec.onnx`, `ocr.res` |-`download_go_deps.py` (§1.4) fetches the five required files — four `.ort` plus-`ocr.res` — into `internal/rag/res/deepdoc/`; `download_deps.py` snapshots the whole-`InfiniFlow/deepdoc` repo and therefore carries both formats.+`download_deps.py` (§1.4) fetches the five required files — four `.ort` plus+`ocr.res` — into `internal/rag/res/deepdoc/`. Python `.onnx` weights require separate provisioning.+The resource image packages these files under+`/huggingface.co/InfiniFlow/deepdoc/` for the Go runtime image.Auto-discovery is **relative to the server process's working directory**:`resolveDeepDocModelDir()` (`cmd/ragflow_server.go`) probes@@ -191,14 +204,14 @@ missing from the model directory.> **Note**: A `internal/rag/res/deepdoc/` populated before the `.ort` switch holds only> `.onnx` and will NOT serve the Go backend, even though the directory looks-> fully populated. Re-run `download_go_deps.py` after updating.+> fully populated. Re-run `download_deps.py` after updating.- **Confirm it is serving** — the server logs, at startup:`in-process DeepDoc backend registered (production backend)`If you instead see a fatal `no in-process DeepDoc backend serving`, it hastwo possible causes: ORT was not linked into the binary, or the modeldirectory is missing one of the five required files listed above. Check the- weights first, then re-run `uv run python3 ragflow_deps/download_go_deps.py` and rebuild+ weights first, then re-run `uv run ragflow_deps/download_deps.py` and rebuild(§1.4 explains the ORT link failure; `build.sh` fails fast with`Error: ONNX Runtime static libraries are not linked` before that happens).diff --git a/internal/ingestion/component/chunker/tool-py/capture_golden.py b/internal/ingestion/component/chunker/tool-py/capture_golden.pyindex aedb2e4a1b..2b3cbe8182 100644--- a/internal/ingestion/component/chunker/tool-py/capture_golden.py+++ b/internal/ingestion/component/chunker/tool-py/capture_golden.py@@ -136,7 +136,7 @@ def _assert_tokenizer_alive() -> None:from rag.nlp import num_tokens_from_stringprobe = "RAGFlow chunker parity tokenizer liveness probe."- hint = "Fix the encoding first: uv run python3 ragflow_deps/download_deps.py (provides ragflow_deps/cl100k_base.tiktoken)."+ hint = "Fix the encoding first: bash build.sh --go provisions ragflow_deps/cl100k_base.tiktoken."# Two shapes to catch. common/token_utils.py builds the encoder on first use and# lets that failure propagate, so an unavailable BPE table raises here; a failure# to encode an otherwise valid string is still reported as 0.diff --git a/internal/ingestion/task/embedder.go b/internal/ingestion/task/embedder.goindex 1fc65c9070..2707130975 100644--- a/internal/ingestion/task/embedder.go+++ b/internal/ingestion/task/embedder.go@@ -96,7 +96,7 @@ func (e *embedder) limiter() tokenizer.Limiter {}if id != "" && !tokenizer.CounterExact(id) {e.limiterErr = fmt.Errorf(- "embedding tokenizer %q is declared for %s but its asset is unavailable (check that ragflow_deps/huggingface.co is present; run `uv run ragflow_deps/download_go_deps.py`): refusing to count with the calibrated estimate",+ "embedding tokenizer %q is declared for %s but its asset is unavailable (check that ragflow_deps/huggingface.co is present; run `uv run ragflow_deps/download_deps.py`): refusing to count with the calibrated estimate",id, e.quotaKey())}e.limiterVal = tokenizer.LimiterFor(id, string(e.quotaKey()), tokenizer.DefaultCalibration())diff --git a/internal/service/nlp/wordnet_helpers_test.go b/internal/service/nlp/wordnet_helpers_test.goindex 9e17f6e377..71a5208b54 100644--- a/internal/service/nlp/wordnet_helpers_test.go+++ b/internal/service/nlp/wordnet_helpers_test.go@@ -21,14 +21,13 @@ import ()// requireWordNetData skips the test when the WordNet dictionary files-// are not present at testWordNetDir. The data ships outside the repo-// (downloaded by Python `download_deps.py`), so CI hosts without it-// must skip rather than fail. The check looks for the lemma-pos offset+// are not present at testWordNetDir. The data ships outside the repo, so+// hosts without it must skip rather than fail. The check looks for the lemma-pos offset// file the loader opens first.func requireWordNetData(t testing.TB) {t.Helper()probe := filepath.Join(testWordNetDir, "index.noun")if _, err := os.Stat(probe); err != nil {- t.Skipf("WordNet data not present at %s; skipping (run download_deps.py to populate)", testWordNetDir)+ t.Skipf("WordNet data not present at %s", testWordNetDir)}}diff --git a/internal/tokenizer/asset_dir_test.go b/internal/tokenizer/asset_dir_test.goindex bcdc6c3e1f..e95e020161 100644--- a/internal/tokenizer/asset_dir_test.go+++ b/internal/tokenizer/asset_dir_test.go@@ -34,7 +34,7 @@ import (const assetDirChildEnv = "RAGFLOW_TOKENIZER_ASSET_DIR_CHILD"-// realSPMAsset is the asset shipped by ragflow_deps/download_go_deps.py, relative to the+// realSPMAsset is the asset shipped by ragflow_deps/download_deps.py, relative to the// repository root.var realSPMAsset = filepath.Join("ragflow_deps", "huggingface.co", "BAAI", "bge-m3", "sentencepiece.bpe.model")@@ -44,7 +44,7 @@ func TestModelAssetsDirIsHonoured(t *testing.T) {}source := filepath.Join("..", "..", realSPMAsset)if _, err := os.Stat(source); err != nil {- t.Skipf("the SPM asset is not present here (%v); run ragflow_deps/download_go_deps.py", err)+ t.Skipf("the SPM asset is not present here (%v); run ragflow_deps/download_deps.py", err)}dir := t.TempDir()diff --git a/internal/tokenizer/bpe_loader.go b/internal/tokenizer/bpe_loader.goindex 17eae01eae..3a4dcc82bd 100644--- a/internal/tokenizer/bpe_loader.go+++ b/internal/tokenizer/bpe_loader.go@@ -56,7 +56,7 @@ func init() {// (223921b76ee99bde995b7ff738513eef100fb51d18c93597a113bcffe865b2a7 for// cl100k_base); that value identifies the path, while the value below verifies// the bytes we actually load. Compute it from the table shipped by-// ragflow_deps/download_go_deps.py: `sha1sum cl100k_base.tiktoken`.+// ragflow_deps/download_deps.py: `sha1sum cl100k_base.tiktoken`.var expectedBpeHashes = map[string]string{"https://openaipublic.blob.core.windows.net/encodings/cl100k_base.tiktoken": "6494e42d5aad2bbb441ea9793af9e7db335c8d9c",}diff --git a/internal/tokenizer/bpe_loader_test.go b/internal/tokenizer/bpe_loader_test.goindex cb00102155..3c3ca84867 100644--- a/internal/tokenizer/bpe_loader_test.go+++ b/internal/tokenizer/bpe_loader_test.go@@ -117,8 +117,8 @@ func TestLocalBpeLoader_ReadsSha1FileFromWorkingDirectory(t *testing.T) {}}-// A developer checkout that has run download_deps.py but never started the-// Python side only has the file under its download name.+// A developer checkout provisioned with build.sh can have the file under+// its download name.func TestLocalBpeLoader_ReadsBundledVocabFromAncestor(t *testing.T) {dir := isolate(t)writeBpeTable(t, filepath.Join(dir, "ragflow_deps", "cl100k_base.tiktoken"), 10)diff --git a/internal/tokenizer/counter_oracle_test.go b/internal/tokenizer/counter_oracle_test.goindex 4ad6d4c019..2e9b22927e 100644--- a/internal/tokenizer/counter_oracle_test.go+++ b/internal/tokenizer/counter_oracle_test.go@@ -118,7 +118,7 @@ func checkCounterAgainstOracle(t *testing.T, fx oracleFixture) {// asset it needs is not on disk.for _, status := range CounterStatuses() {if status.ID == fx.Counter {- t.Fatalf("fixture %q needs counter %q, which is declared but unavailable: its asset is missing from ragflow_deps/huggingface.co (run `uv run ragflow_deps/download_go_deps.py`)", fx.Model, fx.Counter)+ t.Fatalf("fixture %q needs counter %q, which is declared but unavailable: its asset is missing from ragflow_deps/huggingface.co (run `uv run ragflow_deps/download_deps.py`)", fx.Model, fx.Counter)}}t.Fatalf("fixture %q names counter %q, which this build does not implement", fx.Model, fx.Counter)diff --git a/internal/tokenizer/embedding_token_limits.md b/internal/tokenizer/embedding_token_limits.mdindex 9a0ec6c4d9..a9ef2f098e 100644--- a/internal/tokenizer/embedding_token_limits.md+++ b/internal/tokenizer/embedding_token_limits.md@@ -180,7 +180,7 @@ that declares an unavailable counter fails instead of being calibrated silently.| id | family / models | asset (SHA-1 pinned) | implementation ||---|---|---|---|-| `cl100k_base` | OpenAI `text-embedding-3-*`, `ada-002` | `ragflow_deps/cl100k_base.tiktoken` (pre-existing) | tiktoken-go + offline loader (`bpe_loader.go`) |+| `cl100k_base` | OpenAI `text-embedding-3-*`, `ada-002` | `ragflow_deps/cl100k_base.tiktoken` (`download_deps.py`) | tiktoken-go + offline loader (`bpe_loader.go`) || `xlmr-spm` | `bge-m3`, `multilingual-e5-*`, `m3e`, `gte-multilingual`, `jina-embeddings-v3` | `…/BAAI/bge-m3/sentencepiece.bpe.model` `7e88c49f…` | `spm.go`: Unigram Viterbi, byte fallback, NFKC + whitespace + charmap-lite || `bert-wordpiece` | `bge-{large,base,small}-en-v1.5`, `bge-en-icl`, `e5-{base,large}-v2`, `gte-{base,large}`, `jina-embeddings-v2-*` | `…/BAAI/bge-large-en-v1.5/vocab.txt` `c3b41053…` | `wordpiece.go`: BertNormalizer + BertPreTokenizer + `##` longest-match || `qwen-bpe` | `Qwen3-Embedding-*` | `…/Qwen/Qwen3-Embedding-0.6B/tokenizer.json` `e6592f4d…` | `bpe.go`: hand-written Split scanner, GPT-2 byte alphabet, BPE |@@ -347,11 +347,11 @@ an unrelated `vocab.txt` there silently replaces the verified artifact; and audi1. **Identify the architecture** from the model's own artifact (`tokenizer.json`, or`vocab.txt` / `sentencepiece.bpe.model`). If the vocabulary is byte-identical to afamily fixture, only step 4 is needed; if it differs, the model needs its own fixture.-2. **Add the artifact** to `ragflow_deps/download_go_deps.py`'s `TOKENIZER_ASSETS` list -+2. **Add the artifact** to `ragflow_deps/download_deps.py`'s `TOKENIZER_ASSETS` list -individual files, not `snapshot_download` of a multi-GB repo - pin its SHA-1 in theloader map so a mismatch fails at load instead of counting with the wrong table, and add- it to the copy loop in `Dockerfile` / `Dockerfile_base` / `Dockerfile_go` so every runtime image ships it- (the `ragflow_deps` image is built from this script's output, so it needs a refresh too -+ it to the copy loop in `Dockerfile` / `Dockerfile_base` so every runtime image ships it+ (the published `ragflow_deps` image also needs the asset, so refresh it -see §Assets and the oracle for the whole chain).3. **Add the fixture line** to `ORACLES` in `scripts/gen_tokenizer_oracle.py`, generate it,and run `TestCountersMatchOracle`. Passing means every sample agrees, or sits inside the@@ -388,8 +388,7 @@ property tier green, `gofmt` clean.## Assets and the oracle-Assets are fetched, never committed: `ragflow_deps/download_go_deps.py` (the Go-side-downloader - `download_deps.py` is upstream and stays untouched) has a `TOKENIZER_ASSETS`+Assets are fetched, never committed: `ragflow_deps/download_deps.py` has a `TOKENIZER_ASSETS`list that downloads **individual files** (not `snapshot_download` of a repo that alsocarries multi-GB weights) into `ragflow_deps/huggingface.co/<repo>/<file>`, which`.gitignore` already excludes (`huggingface.co/`). Loaders read them from disk with a@@ -397,13 +396,12 @@ SHA-1 pin, walk up from the working and executable directories like `bpe_loader.perform no network I/O. A missing asset fails loudly at every layer that can report one: theruntime image does not build without it (the copy loop below exits non-zero), an embedder whosemodel declares a tokenizer it cannot load refuses to count rather than substituting the-calibrated estimate, and the cl100k table - the one asset that still comes from-`download_deps.py`, which fetches it from the OpenAI blob - has always been a startup panic.+calibrated estimate, and the cl100k table fetched from the OpenAI blob has always been a startup panic when missing.-**The whole chain, because a gap in it is invisible**: `download_go_deps.py` writes-`ragflow_deps/huggingface.co/…`; the `ragflow_deps` image is built from that directory-(`ragflow_deps/Dockerfile` copies `huggingface.co` to `/huggingface.co`); the root-`Dockerfile`, `Dockerfile_base` and `Dockerfile_go` bind-mount that image and copy the four *runtime*+**The whole chain, because a gap in it is invisible**: `download_deps.py` writes+`ragflow_deps/huggingface.co/…`; the published `ragflow_deps` image contains those+files under `/huggingface.co`; the root+`Dockerfile` and `Dockerfile_base` bind-mount that image and copy the four *runtime*tokenizer assets into `/ragflow/ragflow_deps/huggingface.co/…`, which is where thecounters look (the working directory is `/ragflow`). The `tokenizer.json` files thatexist only as cross-check oracles are deliberately **not** shipped. Two consequences@@ -435,11 +433,14 @@ an operator can point at whichever tree they have:| a `huggingface.co`-like root | `<dir>/<repo>/<file>` || a flat directory | `<dir>/<file>` |-The variable is deliberately **not** embedding-specific: DeepDoc's weights use the same-layout (`huggingface.co/InfiniFlow/deepdoc`) and `resolveDeepDocModelDir` honours it too,-so one mount can serve every downloaded model asset. `ragflow_deps/download_go_deps.py` and-the three runtime Dockerfiles keep writing the same tree, so nothing changes for a default-deployment. The layout and the precedence are pinned by tests:+The variable is deliberately **not** embedding-specific: `resolveDeepDocModelDir`+also checks `<dir>/huggingface.co/InfiniFlow/deepdoc` when `MODEL_ASSETS_DIR`+points at `<dir>`. The downloader prepares the five Go DeepDoc files in that tree+under `ragflow_deps/`, as well as in `internal/rag/res/deepdoc/` for local use.+One mounted `ragflow_deps/` directory can therefore serve both asset types.+`ragflow_deps/Dockerfile` packages the tree under `/huggingface.co/` when built+from `ragflow_deps/`.+The layout and the precedence are pinned by tests:`internal/common/model_assets_test.go` for the candidate list, and`internal/tokenizer/asset_dir_test.go` for the end-to-end case (a child process with anunrelated working directory loads the SPM counter from the configured tree).@@ -449,7 +450,7 @@ Two things make the state of the assets observable:- each counter reports the file it loaded (`SourcePath`), so logs and failures can name it;- the ingestor logs an **availability report once at startup** -`embedding tokenizer counters {available: [...]}` plus, when something is missing, a- warning naming the unavailable ones and pointing at `ragflow_deps/download_go_deps.py` /+ warning naming the unavailable ones and pointing at `ragflow_deps/download_deps.py` /`MODEL_ASSETS_DIR`. The report is diagnostics rather than the guard: startup is not fatal(untagged models count with the calibrated estimate by design, and they keep working), butingesting a model that declares a missing asset fails outright - so this report is what@@ -671,8 +672,8 @@ means "what the model does":against the model is not tagged in the catalog; it falls back to the calibratedpath. Guessing a tag is the unsafe direction.2. **Nothing is hand-typed.** Fixtures and expected counts come from the model's- tokenizer; assets come from the downloaders (`download_go_deps.py` for the tokenizer- files, `download_deps.py` for the cl100k table) with a SHA-1 pin.+ tokenizer; model tokenizer assets and the cl100k table come from `download_deps.py`+ with a SHA-1 pin.3. **Equality, not resemblance** (see above), and the dangerous direction(under-count) is the one the tests are shaped around.4. **Every approximation is named** in §Known approximations together with thediff --git a/internal/tokenizer/embedlimit.go b/internal/tokenizer/embedlimit.goindex b6990df1ba..18b56075d0 100644--- a/internal/tokenizer/embedlimit.go+++ b/internal/tokenizer/embedlimit.go@@ -663,7 +663,7 @@ func RefuseUnavailableCounter(tokenizerID, calibrationKey string) error {return nil}return fmt.Errorf(- "model tokenizer %q is declared for %s but its asset is unavailable (check that ragflow_deps/huggingface.co is present; run `uv run ragflow_deps/download_go_deps.py`): refusing to count with the calibrated estimate",+ "model tokenizer %q is declared for %s but its asset is unavailable (check that ragflow_deps/huggingface.co is present; run `uv run ragflow_deps/download_deps.py`): refusing to count with the calibrated estimate",tokenizerID, calibrationKey)}diff --git a/internal/tokenizer/spm.go b/internal/tokenizer/spm.goindex 03540bcfc5..2e88499f8d 100644--- a/internal/tokenizer/spm.go+++ b/internal/tokenizer/spm.go@@ -850,7 +850,7 @@ func readTokenizerAsset(names []string, pinKey string, pins map[string]string) (}return candidate, raw, nil}- return "", nil, fmt.Errorf("no tokenizer asset found; run `uv run ragflow_deps/download_go_deps.py` or set %s to a directory holding them; tried: %s",+ return "", nil, fmt.Errorf("no tokenizer asset found; run `uv run ragflow_deps/download_deps.py` or set %s to a directory holding them; tried: %s",common.EnvModelAssetsDir, strings.Join(tried, ", "))}diff --git a/ragflow_deps/Dockerfile b/ragflow_deps/Dockerfileindex 236b38eb71..ade16bfd9b 100644--- a/ragflow_deps/Dockerfile+++ b/ragflow_deps/Dockerfile@@ -1,10 +1,7 @@-# This builds an image that contains the resources needed by Dockerfile-#+# Build from ragflow_deps/ after running download_deps.py:+# docker build -f Dockerfile -t infiniflow/ragflow_deps:latest .FROM scratch-# Copy resources downloaded via download_deps.py-COPY chromedriver-linux64-121-0-6167-85 chrome-linux64-121-0-6167-85 cl100k_base.tiktoken libssl1.1_1.1.1f-1ubuntu2_amd64.deb libssl1.1_1.1.1f-1ubuntu2_arm64.deb tika-server-standard-3.3.0.jar tika-server-standard-3.3.0.jar.md5 libssl*.deb uv-x86_64-unknown-linux-gnu.tar.gz uv-aarch64-unknown-linux-gnu.tar.gz stagehand-server-v3-linux-x64 stagehand-server-v3-linux-arm64 /--COPY nltk_data /nltk_data+COPY cl100k_base.tiktoken stagehand-server-v3-linux-x64 stagehand-server-v3-linux-arm64 /COPY huggingface.co /huggingface.codiff --git a/ragflow_deps/download_deps.py b/ragflow_deps/download_deps.pyindex fdb8ad2536..0a7f3970ab 100755--- a/ragflow_deps/download_deps.py+++ b/ragflow_deps/download_deps.py@@ -4,63 +4,71 @@# /// script# requires-python = ">=3.10"# dependencies = [-# "nltk",-# "huggingface-hub"+# "huggingface-hub",+# "requests",# ]# ///-# This script downloads every artifact that the `infiniflow/ragflow_deps`-# Docker image bakes in. Run it from anywhere — the `__main__` block-# chdir's into this file's own directory, so all outputs land under-# `ragflow_deps/` regardless of the caller's CWD.+# This script prepares Go dependencies on Linux x86_64: the native static+# libraries (pdfium / pdf_oxide / office_oxide / onnxruntime) for `build.sh`,+# and the Go DeepDoc `.ort` weights (det/layout/tsr/rec.ort + ocr.res) so a Go+# dev can run the in-process backend locally. It also downloads the cl100k BPE+# table and installs the local stagehand driver in the Go SDK's cache.+# Archives and tokenizer assets land under ragflow_deps/; DeepDoc weights land+# under internal/rag/res/deepdoc/, regardless of the caller's working directory.#-# Build-context relationship: `ragflow_deps/Dockerfile` is built with-# `ragflow_deps/` as its build context, so the files written here MUST-# sit at the top of `ragflow_deps/`. The Dockerfile's COPY lines assume-# top-level paths (`huggingface.co`, `nltk_data`, `cl100k_base.tiktoken`,-# `*.deb`, `*.jar`, `*.tar.gz`, `stagehand-server-v3-linux-<arch>`).+# Downloaded archives and tokenizer assets land under `ragflow_deps/`.## Typical workflow:#-# uv run python3 ragflow_deps/download_deps.py # download-# cd ragflow_deps-# docker build -f Dockerfile -t infiniflow/ragflow_deps .+# uv run ragflow_deps/download_deps.py+# Go DeepDoc weights: in addition to the native libs, this script downloads the+# five Go model files (internal/common.DeepDocModelFiles) from InfiniFlow/deepdoc+# straight into the repo's canonical model directory `internal/rag/res/deepdoc/` (one level+# up from this script). The Go server auto-discovers that directory via+# resolveDeepDocModelDir(), and build.sh --test-native / internal/deepdoc/native/run.sh+# default MODEL_DIR there too — so after running this script NO MODEL_DIR env needs+# to be set:#-# The main `Dockerfile` (built from the project root) pulls this image-# via `--mount=type=bind,from=infiniflow/ragflow_deps:latest,...` and-# is unaffected by where these files live locally.+# bash build.sh --test-native # or: cd internal/deepdoc/native && bash run.sh+#+# Embedding tokenizer assets: the Go token counters in internal/tokenizer load a+# tokenizer file per family (XLM-R SentencePiece, BERT WordPiece, two byte-level BPE+# families) from `huggingface.co/<repo>/<file>` at the top of ragflow_deps/ - the path+# the loaders search and the one ragflow_deps/Dockerfile ships. This script fetches+# them too. Models that declare a missing tokenizer cannot be ingested+# (see internal/tokenizer/embedding_token_limits.md).import argparseimport hashlibimport os+import platform+import reimport shutilimport sys-import urllib.request--# NLTK >=3.10 refuses proxied downloads (SSRF guard) unless opted in; the-# runners sit behind a proxy, so allow proxied fetches before importing nltk.-os.environ.setdefault("NLTK_ALLOW_PROXIED_URLOPEN", "1")--import nltk-from huggingface_hub import snapshot_download--# mirrors internal/common.DeepDocORTVersion (Go in-process backend). ONE OF-# FOUR places (with that Go constant, ORT_VERSION in ragflow_deps/download_go_deps.py,-# and ARG ORT_VERSION in Dockerfile_go) that must carry the same ONNX Runtime-# native release for the statically-linked Go DeepDoc backend. This file's-# download URL and extracted dir name are derived from ORT_VERSION here, but-# there is no single source of truth — keep all four equal. build.sh-# --check-ort-version greps this file (and the other three) to fail fast on-# drift. (The Python pip onnxruntime== pin in pyproject.toml is versioned-# independently and is not part of this check.)+import tarfile+import tempfile+import zipfile++import requests++# Mirrors internal/common.DeepDocORTVersion (Go in-process backend). ONE OF+# THREE places (with that Go constant and ARG ORT_VERSION in Dockerfile)+# that must carry the same ONNX Runtime+# native release for the statically-linked Go DeepDoc backend. There is no+# single source of truth — keep all three equal. build.sh --check-ort-version+# greps this file (and the other two) to fail fast on drift. (The Python pip+# onnxruntime== pin in pyproject.toml is versioned independently and is not+# part of this check.)## Source of the native static archives: infiniflow/ragflow-build (our own# ORT-only minimal build), NOT the third-party csukuangfj/onnxruntime-libs# account. The release tag is `onnxruntime-v{ORT_VERSION}` and the asset is# `onnxruntime-v{ORT_VERSION}-linux-x86_64.zip`. The archive is occasionally-# re-issued under this SAME tag/asset name with patched content; download_deps.py-# detects that via a `{asset}.sha256` sidecar and re-downloads/re-extracts, so a-# stale local copy never silently lingers.+# re-issued under this SAME tag/asset name with patched content; this script+# detects that via a `{asset}.sha256`+# sidecar and re-download/re-extract, so a stale local copy never silently+# lingers.ORT_VERSION = "1.29.0"@@ -88,18 +96,115 @@ def _sha256_of(path):return h.hexdigest()+# Mirrors internal/common.DeepDocModelFiles (Go in-process DeepDoc backend).+# These are the ONLY weights the Go backend loads; the full InfiniFlow/deepdoc+# repo also ships .onnx (Python-only), which this Go-only script deliberately+# skips to keep the download lean.+DEEPDOC_REPO = "InfiniFlow/deepdoc"+DEEPDOC_MODEL_FILES = ["det.ort", "layout.ort", "tsr.ort", "rec.ort", "ocr.res"]++# Embedding tokenizer assets for the Go counters (internal/tokenizer). Each entry is+# (repo, file, kind):+#+# "runtime" - the counters load it in production, so the runtime images must ship it+# (the copy loop in Dockerfile / Dockerfile_base) and the loader must pin+# its SHA-1;+# "oracle" - only scripts/gen_tokenizer_oracle.py needs it, to regenerate the test+# fixtures, so it is deliberately NOT shipped to the image.+#+# ragflow_deps/test_tokenizer_assets.py checks these download, packaging, and pin lists.+#+# Fetched per file, not by snapshot: these repos also carry multi-GB weights we do not+# want. Missing tokenizer assets fail dependency preparation.+TOKENIZER_ASSETS = [+ # XLM-R SentencePiece (Unigram) - the BAAI bge / multilingual-e5 / m3e /+ # gte-multilingual / jina-v3 family shares this vocabulary.+ ("BAAI/bge-m3", "sentencepiece.bpe.model", "runtime"),+ # BERT WordPiece vocab for the bge-*-en / e5-* / gte-base family.+ ("BAAI/bge-large-en-v1.5", "vocab.txt", "runtime"),+ # Byte-level BPE families (Qwen3-Embedding, Mistral/Llama embeddings).+ ("Qwen/Qwen3-Embedding-0.6B", "tokenizer.json", "runtime"),+ ("intfloat/e5-mistral-7b-instruct", "tokenizer.json", "runtime"),+ # Cross-check oracles only: the HF tokenizer.json files the same families convert to.+ ("BAAI/bge-m3", "tokenizer.json", "oracle"),+ ("BAAI/bge-large-en-v1.5", "tokenizer.json", "oracle"),+]+++def prune_stale_onnxruntime(static_lib_dir, version):+ """Remove ONNX Runtime version dirs under static_lib that do NOT match+ `version`. Without this, a version bump leaves the stale dir next to+ the new one and build.sh's `find ... -name '*.a'` links BOTH (duplicate+ symbols / wrong version, silently)."""+ if not os.path.isdir(static_lib_dir):+ return+ expected = _ort_normalized_dir(version)+ for name in os.listdir(static_lib_dir):+ if not name.startswith("onnxruntime-linux-x64-static_lib-"):+ continue+ if name == expected:+ continue+ stale = os.path.join(static_lib_dir, name)+ print(f" Removing stale ONNX Runtime dir: {stale}")+ shutil.rmtree(stale)+++def has_static_archives(directory):+ """True when `directory` holds at least one static archive (.a)."""+ return any(f.endswith(".a") for _, _, files in os.walk(directory) for f in files)+++def extract_onnxruntime(static_lib_dir, archive_path, version):+ """Ensure the ONNX Runtime static archives for `version` sit under+ `static_lib_dir`. Returns True when that version is available afterwards+ (extracted now or already present), False when the archive is missing.++ The infiniflow/ragflow-build release zip carries a top-level dir named+ onnxruntime-v{version}-linux-x86_64, so a present+ `static_lib_dir` is NOT evidence that THIS version is extracted: after a+ version bump the stale dir is pruned and the new one must be extracted.+ """+ if not os.path.isfile(archive_path):+ print(f" Skipping extraction: {os.path.basename(archive_path)} not found")+ return False+ prune_stale_onnxruntime(static_lib_dir, version)+ version_dir = os.path.join(static_lib_dir, _ort_normalized_dir(version))+ if os.path.isdir(version_dir) and has_static_archives(version_dir):+ print(f" ✓ onnxruntime/static_lib ({version}) already extracted to {version_dir}")+ return True+ os.makedirs(static_lib_dir, exist_ok=True)+ print(f" Extracting {os.path.basename(archive_path)} → {static_lib_dir}")+ with zipfile.ZipFile(archive_path) as zf:+ zf.extractall(static_lib_dir)+ # The infiniflow/ragflow-build release zip carries a top-level dir named+ # onnxruntime-v{version}-linux-x86_64, but build.sh's glob and the stale+ # checks above all expect onnxruntime-linux-x64-static_lib-{version}-glibc2_28.+ # Rename it so every consumer shares one name convention (driven by+ # ORT_VERSION).+ extracted = os.path.join(static_lib_dir, _ort_extracted_dir(version))+ normalized = os.path.join(static_lib_dir, _ort_normalized_dir(version))+ if os.path.isdir(extracted) and extracted != normalized:+ if os.path.exists(normalized):+ shutil.rmtree(normalized)+ print(f" Renaming {os.path.basename(extracted)} → {os.path.basename(normalized)}")+ os.rename(extracted, normalized)+ return True+++def _go_module_version(module):+ repo_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))+ with open(os.path.join(repo_root, "go.mod")) as file:+ version = re.search(rf"{re.escape(module)}\s+v([^\s]+)", file.read())+ if version is None:+ raise RuntimeError(f"{module} version not found in go.mod")+ return version.group(1)++def get_urls(use_china_mirrors=False) -> list[str | list[str]]:+ office_version = _go_module_version("github.com/yfedoseev/office_oxide/go")if use_china_mirrors:return [- "http://mirrors.tuna.tsinghua.edu.cn/ubuntu/pool/main/o/openssl/libssl1.1_1.1.1f-1ubuntu2_amd64.deb",- "http://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/pool/main/o/openssl/libssl1.1_1.1.1f-1ubuntu2_arm64.deb",- "https://repo.huaweicloud.com/repository/maven/org/apache/tika/tika-server-standard/3.3.0/tika-server-standard-3.3.0.jar",- "https://repo.huaweicloud.com/repository/maven/org/apache/tika/tika-server-standard/3.3.0/tika-server-standard-3.3.0.jar.md5","https://openaipublic.blob.core.windows.net/encodings/cl100k_base.tiktoken",- ["https://registry.npmmirror.com/-/binary/chrome-for-testing/121.0.6167.85/linux64/chrome-linux64.zip", "chrome-linux64-121-0-6167-85"],- ["https://registry.npmmirror.com/-/binary/chrome-for-testing/121.0.6167.85/linux64/chromedriver-linux64.zip", "chromedriver-linux64-121-0-6167-85"],- "https://github.com/astral-sh/uv/releases/download/0.9.16/uv-x86_64-unknown-linux-gnu.tar.gz",- "https://github.com/astral-sh/uv/releases/download/0.9.16/uv-aarch64-unknown-linux-gnu.tar.gz",# stagehand-server-v3 Node.js SEA binaries (used by Browser# component in local mode).#@@ -114,42 +219,22 @@ def get_urls(use_china_mirrors=False) -> list[str | list[str]]:# latest server release. There is no version# correspondence to maintain; "both on v3.x" is the# compatibility contract.- "https://github.com/browserbase/stagehand/releases/download/stagehand-server-v3/v3.7.2/stagehand-server-v3-linux-x64",- "https://github.com/browserbase/stagehand/releases/download/stagehand-server-v3/v3.7.2/stagehand-server-v3-linux-arm64",- # Native static libraries for Go build (pdfium, pdf_oxide, office_oxide)- # Used by build.sh's check_*_deps functions — pre-downloaded to avoid- # network access during CI.- ["https://github.com/kognitos/pdfium-static/releases/download/chromium%2F7809/pdfium-linux-x64-static.tgz", "pdfium-linux-x64-static.tgz"],- ["https://github.com/yfedoseev/pdf_oxide/releases/download/v0.3.73/pdf_oxide-go-ffi-linux-amd64.tar.gz", "pdf_oxide-go-ffi-linux-amd64.tar.gz"],- ["https://github.com/yfedoseev/office_oxide/releases/download/v0.1.12/native-linux-x86_64.tar.gz", "office_oxide-linux-x86_64.tar.gz"],- # ONNX Runtime static archives for the Go in-process (DeepDoc)- # backend. Statically linked into the server binary (see build.sh:- # ONNXRUNTIME_STATIC_PREFIX — no --whole-archive, so unreferenced- # kernels are dropped; only OrtGetApiBase is exported, via- # --dynamic-list), so no libonnxruntime.so is needed at runtime —- # OrtGetApiBase is resolved via dlopen(NULL) (the process-global- # symbol table, not the executable's own path). Our own- # infiniflow/ragflow-build ORT-only minimal build- # (onnxruntime-v{ORT_VERSION}) is CPU-only and glibc2_28-based,- # matching ORT_VERSION's C-API line (ABI-compatible with- # onnxruntime_go) and the onnxruntime the Python goldens were- # generated with.+ "https://gh-proxy.com/https://github.com/browserbase/stagehand/releases/download/stagehand-server-v3/v3.7.2/stagehand-server-v3-linux-x64",+ "https://gh-proxy.com/https://github.com/browserbase/stagehand/releases/download/stagehand-server-v3/v3.7.2/stagehand-server-v3-linux-arm64",+ # Native static libraries for Go build (pdfium, pdf_oxide,+ # office_oxide, onnxruntime). Used by build.sh's check_*_deps+ # functions — pre-downloaded to avoid network access during CI.+ ["https://gh-proxy.com/https://github.com/kognitos/pdfium-static/releases/download/chromium%2F7809/pdfium-linux-x64-static.tgz", "pdfium-linux-x64-static.tgz"],+ ["https://gh-proxy.com/https://github.com/yfedoseev/pdf_oxide/releases/download/v0.3.73/pdf_oxide-go-ffi-linux-amd64.tar.gz", "pdf_oxide-go-ffi-linux-amd64.tar.gz"],+ [f"https://gh-proxy.com/https://github.com/yfedoseev/office_oxide/releases/download/v{office_version}/native-linux-x86_64.tar.gz", "office_oxide-linux-x86_64.tar.gz"],[- f"https://github.com/infiniflow/ragflow-build/releases/download/onnxruntime-v{ORT_VERSION}/{_ort_asset_name(ORT_VERSION)}",+ f"https://gh-proxy.com/https://github.com/infiniflow/ragflow-build/releases/download/onnxruntime-v{ORT_VERSION}/{_ort_asset_name(ORT_VERSION)}",_ort_asset_name(ORT_VERSION),],]else:return [- "http://archive.ubuntu.com/ubuntu/pool/main/o/openssl/libssl1.1_1.1.1f-1ubuntu2_amd64.deb",- "http://ports.ubuntu.com/pool/main/o/openssl/libssl1.1_1.1.1f-1ubuntu2_arm64.deb",- "https://repo1.maven.org/maven2/org/apache/tika/tika-server-standard/3.3.0/tika-server-standard-3.3.0.jar",- "https://repo1.maven.org/maven2/org/apache/tika/tika-server-standard/3.3.0/tika-server-standard-3.3.0.jar.md5","https://openaipublic.blob.core.windows.net/encodings/cl100k_base.tiktoken",- ["https://storage.googleapis.com/chrome-for-testing-public/121.0.6167.85/linux64/chrome-linux64.zip", "chrome-linux64-121-0-6167-85"],- ["https://storage.googleapis.com/chrome-for-testing-public/121.0.6167.85/linux64/chromedriver-linux64.zip", "chromedriver-linux64-121-0-6167-85"],- "https://github.com/astral-sh/uv/releases/download/0.9.16/uv-x86_64-unknown-linux-gnu.tar.gz",- "https://github.com/astral-sh/uv/releases/download/0.9.16/uv-aarch64-unknown-linux-gnu.tar.gz",# stagehand-server-v3 Node.js SEA binaries (used by Browser# component in local mode).#@@ -166,24 +251,12 @@ def get_urls(use_china_mirrors=False) -> list[str | list[str]]:# compatibility contract."https://github.com/browserbase/stagehand/releases/download/stagehand-server-v3/v3.7.2/stagehand-server-v3-linux-x64","https://github.com/browserbase/stagehand/releases/download/stagehand-server-v3/v3.7.2/stagehand-server-v3-linux-arm64",- # Native static libraries for Go build (pdfium, pdf_oxide, office_oxide)- # Used by build.sh's check_*_deps functions — pre-downloaded to avoid- # network access during CI.+ # Native static libraries for Go build (pdfium, pdf_oxide,+ # office_oxide, onnxruntime). Used by build.sh's check_*_deps+ # functions — pre-downloaded to avoid network access during CI.["https://github.com/kognitos/pdfium-static/releases/download/chromium%2F7809/pdfium-linux-x64-static.tgz", "pdfium-linux-x64-static.tgz"],["https://github.com/yfedoseev/pdf_oxide/releases/download/v0.3.73/pdf_oxide-go-ffi-linux-amd64.tar.gz", "pdf_oxide-go-ffi-linux-amd64.tar.gz"],- ["https://github.com/yfedoseev/office_oxide/releases/download/v0.1.12/native-linux-x86_64.tar.gz", "office_oxide-linux-x86_64.tar.gz"],- # ONNX Runtime static archives for the Go in-process (DeepDoc)- # backend. Statically linked into the server binary (see build.sh:- # ONNXRUNTIME_STATIC_PREFIX — no --whole-archive, so unreferenced- # kernels are dropped; only OrtGetApiBase is exported, via- # --dynamic-list), so no libonnxruntime.so is needed at runtime —- # OrtGetApiBase is resolved via dlopen(NULL) (the process-global- # symbol table, not the executable's own path). Our own- # infiniflow/ragflow-build ORT-only minimal build- # (onnxruntime-v{ORT_VERSION}) is CPU-only and glibc2_28-based,- # matching ORT_VERSION's C-API line (ABI-compatible with- # onnxruntime_go) and the onnxruntime the Python goldens were- # generated with.+ [f"https://github.com/yfedoseev/office_oxide/releases/download/v{office_version}/native-linux-x86_64.tar.gz", "office_oxide-linux-x86_64.tar.gz"],[f"https://github.com/infiniflow/ragflow-build/releases/download/onnxruntime-v{ORT_VERSION}/{_ort_asset_name(ORT_VERSION)}",_ort_asset_name(ORT_VERSION),@@ -191,23 +264,238 @@ def get_urls(use_china_mirrors=False) -> list[str | list[str]]:]-repos = [- "InfiniFlow/text_concat_xgb_v1.0",- "InfiniFlow/deepdoc",-]+def download_with_progress(url, filename):+ filename = os.fspath(filename)+ temporary = None+ try:+ with requests.get(url, stream=True, timeout=(15, 60)) as response:+ response.raise_for_status()+ total_size = int(response.headers.get("content-length", 0))+ with tempfile.NamedTemporaryFile(dir=os.path.dirname(os.path.abspath(filename)), delete=False) as file:+ temporary = file.name+ downloaded = 0+ for data in response.iter_content(1 << 20):+ file.write(data)+ downloaded += len(data)+ if total_size > 0:+ progress = (downloaded / total_size) * 100+ sys.stdout.write(f"\rProgress: {progress:.1f}% ({downloaded}/{total_size} bytes)")+ sys.stdout.flush()+ if downloaded == 0:+ raise requests.RequestException(f"Empty dependency download: {url}")+ os.replace(temporary, filename)+ finally:+ if temporary is not None and os.path.exists(temporary):+ os.unlink(temporary)+ print()+++def _extract_native_archive(archive_path, target, required_files, version=None):+ def complete(directory):+ for name in required_files:+ path = os.path.join(directory, name)+ if not os.path.isfile(path) or os.path.getsize(path) == 0:+ return False+ if version is not None and name.endswith(".a"):+ with open(path, "rb") as file:+ if b"\x00" + version.encode() + b"\x00" not in file.read():+ return False+ return True++ if complete(target):+ print(f" ✓ {os.path.basename(target)} already extracted to {target}")+ return+ os.makedirs(os.path.dirname(target), exist_ok=True)+ with tempfile.TemporaryDirectory(dir=os.path.dirname(target)) as staging:+ with tarfile.open(archive_path) as archive:+ members = archive.getmembers()+ for member in members:+ if os.path.isabs(member.name) or ".." in member.name.split("/") or not (member.isfile() or member.isdir()):+ raise tarfile.TarError(f"Unsafe native archive member: {member.name}")+ archive.extractall(staging, members=members)+ if not complete(staging):+ raise RuntimeError(f"{archive_path} is missing required native files or version {version}")+ if os.path.isdir(target):+ shutil.rmtree(target)+ os.replace(staging, target)+ print(f" ✓ extracted {os.path.basename(archive_path)} to {target}")+++def _install_stagehand():+ if platform.system() != "Linux":+ raise RuntimeError("The downloaded native dependencies support Linux only")+ architecture = {"x86_64": "x64", "aarch64": "arm64"}.get(platform.machine())+ if architecture is None:+ raise RuntimeError(f"Unsupported stagehand architecture: {platform.machine()}")+ version = _go_module_version("github.com/browserbase/stagehand-go/v3")+ cache_root = os.environ.get("XDG_CACHE_HOME") or os.path.expanduser("~/.cache")+ target_dir = os.path.join(cache_root, "stagehand", "lib", f"go_{version}")+ os.makedirs(target_dir, exist_ok=True)+ filename = f"stagehand-server-v3-linux-{architecture}"+ source = os.path.join(os.path.dirname(os.path.abspath(__file__)), filename)+ target = os.path.join(target_dir, filename)+ with tempfile.TemporaryDirectory(dir=target_dir) as staging:+ prepared = os.path.join(staging, filename)+ shutil.copyfile(source, prepared)+ os.chmod(prepared, 0o755)+ os.replace(prepared, target)+ print(f" ✓ stagehand installed to {target}")+++def _valid_download(filename):+ if not os.path.isfile(filename) or os.path.getsize(filename) == 0:+ return False+ if filename.endswith((".tar.gz", ".tgz")):+ if not tarfile.is_tarfile(filename):+ return False+ if os.path.basename(filename) == "office_oxide-linux-x86_64.tar.gz":+ version = _go_module_version("github.com/yfedoseev/office_oxide/go")+ with tarfile.open(filename) as archive:+ try:+ member = next((member for member in archive.getmembers() if os.path.normpath(member.name) == "lib/liboffice_oxide.a"), None)+ if member is None:+ return False+ library = archive.extractfile(member)+ if library is None:+ return False+ with library:+ return b"\x00" + version.encode() + b"\x00" in library.read()+ except (KeyError, tarfile.TarError, EOFError):+ return False+ return True+ if filename.endswith(".zip"):+ return zipfile.is_zipfile(filename)+ if os.path.basename(filename).startswith("stagehand-server-"):+ with open(filename, "rb") as file:+ return file.read(4) == b"\x7fELF"+ return True+++def download_go_models(use_china_mirrors=False):+ """Download the Go DeepDoc `.ort` weights so a Go dev can run the in-process+ backend with no further setup.++ The files are written into the repo's canonical model directory+ `internal/rag/res/deepdoc/` (relative to the repo root, one level up from this+ script). The Go server auto-discovers that directory via+ resolveDeepDocModelDir(), and build.sh --test-native / run.sh default+ MODEL_DIR there too — so after this script runs, no MODEL_DIR env needs to+ be set.++ Uses hf_hub_download per-file (not snapshot_download) to fetch only the+ five Go model files; the `.onnx` siblings are Python-only and skipped.+ On China mirrors, route through hf-mirror.com via HF_ENDPOINT.+ """+ if use_china_mirrors:+ os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"+ # Imported lazily so the module stays importable (and its unit tests stay+ # huggingface-free) without the huggingface_hub dependency installed.+ from huggingface_hub import hf_hub_download++ # Canonical local-dev model dir the Go backend auto-discovers.+ repo_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))+ target_dir = os.path.join(repo_root, "internal", "rag", "res", "deepdoc")+ os.makedirs(target_dir, exist_ok=True)++ missing = []+ for fname in DEEPDOC_MODEL_FILES:+ dest = os.path.join(target_dir, fname)+ if os.path.isfile(dest) and os.path.getsize(dest) > 0:+ print(f" ✓ {fname} already present")+ continue+ print(f"Downloading deepdoc model {fname}...")+ try:+ hf_hub_download(repo_id=DEEPDOC_REPO, filename=fname, local_dir=target_dir)+ if not os.path.isfile(dest) or os.path.getsize(dest) == 0:+ raise RuntimeError(f"Empty or missing model: {dest}")+ except Exception as e: # noqa: BLE001 - collected and surfaced below+ missing.append((fname, e))++ if missing:+ for fname, e in missing:+ print(f" ERROR: failed to download {fname}: {e}", file=sys.stderr)+ print(+ "\n"+ "The Go in-process DeepDoc backend loads ONLY the .ort weights listed in\n"+ "internal/common.DeepDocModelFiles. They are fetched from "+ f"{DEEPDOC_REPO} into:\n"+ f" {target_dir}\n"+ "Without them the backend cannot serve — the server exits with a fatal\n"+ '"no in-process DeepDoc backend serving". To recover:\n'+ " - re-run this script (a transient HF/network error usually clears);\n"+ " - behind the GFW, re-run with --china-mirrors (routes via hf-mirror.com);\n"+ " - or copy the missing files into that directory by hand.",+ file=sys.stderr,+ )+ sys.exit(1)++ # Keep the dependency image's existing huggingface.co build-context layout.+ image_dir = os.path.join(repo_root, "ragflow_deps", "huggingface.co", DEEPDOC_REPO)+ os.makedirs(image_dir, exist_ok=True)+ for fname in DEEPDOC_MODEL_FILES:+ shutil.copyfile(os.path.join(target_dir, fname), os.path.join(image_dir, fname))+ print(f" ✓ Go DeepDoc models ready under {target_dir}")+ print(" No MODEL_DIR env needed: the Go backend auto-discovers this directory.")-def download_model(repository_id):- local_directory = os.path.abspath(os.path.join("huggingface.co", repository_id))- os.makedirs(local_directory, exist_ok=True)- snapshot_download(repo_id=repository_id, local_dir=local_directory)++def download_tokenizer_assets(use_china_mirrors=False):+ """Download the embedding tokenizer assets the Go counters load.++ Written to `ragflow_deps/huggingface.co/<repo>/<file>`: the path internal/tokenizer+ searches and the one ragflow_deps/Dockerfile builds its image from. Fetched per file+ (these repos also carry multi-GB weights we do not want), routed via hf-mirror.com+ when --china-mirrors is set.++ A missing asset is fatal: models declaring its tokenizer refuse ingestion.+ Runtime images also reject missing files during their copy step.+ """+ if use_china_mirrors:+ os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"+ # Imported lazily so the module stays importable (and its unit tests stay+ # huggingface-free) without the huggingface_hub dependency installed.+ from huggingface_hub import hf_hub_download++ base = os.path.join(os.path.dirname(os.path.abspath(__file__)), "huggingface.co")+ missing = []+ for repo_id, filename, _kind in TOKENIZER_ASSETS:+ target_dir = os.path.join(base, repo_id)+ target_file = os.path.join(target_dir, filename)+ if os.path.isfile(target_file) and os.path.getsize(target_file) > 0:+ print(f" ✓ {repo_id}/{filename} already present")+ continue+ os.makedirs(target_dir, exist_ok=True)+ print(f"Downloading tokenizer asset {repo_id}/{filename}...")+ try:+ hf_hub_download(repo_id=repo_id, filename=filename, local_dir=target_dir)+ if not os.path.isfile(target_file) or os.path.getsize(target_file) == 0:+ raise RuntimeError(f"Empty or missing tokenizer asset: {target_file}")+ except Exception as e: # noqa: BLE001 - collected and surfaced below+ missing.append((repo_id, filename, e))++ if missing:+ for repo_id, filename, e in missing:+ print(f" ERROR: could not fetch {repo_id}/{filename}: {e}", file=sys.stderr)+ print(+ "\n"+ "The embedding tokenizer counters in internal/tokenizer load these files from\n"+ f" {base}\n"+ "Without one, models that declare its tokenizer cannot be ingested.\n"+ "The runtime images refuse to\n"+ "build on the same condition (their copy loops exit 1), so this script fails\n"+ "instead of handing a broken tree to the image build. To recover:\n"+ " - re-run this script (a transient HF/network error usually clears);\n"+ " - behind the GFW, re-run with --china-mirrors (routes via hf-mirror.com);\n"+ " - or point MODEL_ASSETS_DIR at a directory holding the same layout.\n",+ file=sys.stderr,+ )+ sys.exit(1)++ print(" ✓ embedding tokenizer assets ready")if __name__ == "__main__":- # Anchor CWD to this file's directory so all relative outputs- # (huggingface.co/, nltk_data/, *.deb, *.jar, *.tar.gz, etc.) land- # at the top of ragflow_deps/ regardless of where the user invokes- # the script from. This is the build context for `ragflow_deps/Dockerfile`.+ # Anchor archives, the BPE table, and tokenizer assets to ragflow_deps/.os.chdir(os.path.dirname(os.path.abspath(__file__)))parser = argparse.ArgumentParser(description="Download dependencies with optional China mirror support")@@ -216,12 +504,6 @@ if __name__ == "__main__":urls = get_urls(args.china_mirrors)- # Some mirrors (e.g. archive.ubuntu.com) reject the default urllib- # User-Agent with HTTP 403, so install an opener with a browser-like UA.- opener = urllib.request.build_opener()- opener.addheaders = [("User-Agent", "Mozilla/5.0")]- urllib.request.install_opener(opener)-for url in urls:download_url = url[0] if isinstance(url, list) else urlfilename = url[1] if isinstance(url, list) else url.split("/")[-1]@@ -232,20 +514,20 @@ if __name__ == "__main__":# exports SessionGetInitializer*). A pure existence check would then# keep a colleague's stale local copy and fail to link onnxruntime_go.# Verify against the published .sha256 sidecar so a re-issued archive- # is always re-downloaded and re-extracted. Every other archive keeps- # the legacy existence-based skip.+ # is always re-downloaded and re-extracted.is_ort = filename == _ort_asset_name(ORT_VERSION)expected_sha = Noneif is_ort:sidecar_url = download_url + ".sha256"try:- with urllib.request.urlopen(sidecar_url) as resp:- expected_sha = resp.read().decode().split()[0]+ resp = requests.get(sidecar_url, timeout=30)+ resp.raise_for_status()+ expected_sha = resp.text.split()[0]except Exception as exc: # noqa: BLE001 - best-effort; fall back to legacyprint(f" WARNING: could not fetch {sidecar_url} ({exc}); skipping checksum for {filename}")needs_download = True- if os.path.exists(filename):+ if _valid_download(filename):if expected_sha is not None:actual = _sha256_of(filename)if actual == expected_sha:@@ -257,13 +539,17 @@ if __name__ == "__main__":needs_download = Falseif needs_download:- urllib.request.urlretrieve(download_url, filename)+ download_with_progress(download_url, filename)if expected_sha is not None:actual = _sha256_of(filename)if actual != expected_sha:+ os.unlink(filename)print(f" ERROR: {filename} checksum mismatch after download (got {actual}, expected {expected_sha})", file=sys.stderr)sys.exit(1)print(f" ✓ {filename} checksum verified ({actual})")+ if not _valid_download(filename):+ os.unlink(filename)+ raise RuntimeError(f"Invalid downloaded dependency: {filename}")# Force re-extract below: drop any previously extracted version dir# so the same-named re-issued archive actually refreshes the .a files.if is_ort:@@ -276,77 +562,43 @@ if __name__ == "__main__":# Extract native static libraries to ~/ragflow-native-libs for Go build.# Ensures build.sh can find them without network access.native_deps_dir = os.path.expanduser("~/ragflow-native-libs")+ office_version = _go_module_version("github.com/yfedoseev/office_oxide/go")extractions = [- ("pdfium-linux-x64-static.tgz", "pdfium-static"),- ("pdf_oxide-go-ffi-linux-amd64.tar.gz", "pdf_oxide"),- ("office_oxide-linux-x86_64.tar.gz", "office_oxide"),- (_ort_asset_name(ORT_VERSION), os.path.join("onnxruntime", "static_lib")),+ ("pdfium-linux-x64-static.tgz", "pdfium-static", ["lib/libpdfium.a", "lib/libc++.a", "lib/libc++abi.a", "include/fpdfview.h"], None),+ ("pdf_oxide-go-ffi-linux-amd64.tar.gz", "pdf_oxide", ["lib/linux_amd64/libpdf_oxide.a", "include/pdf_oxide.h"], None),+ ("office_oxide-linux-x86_64.tar.gz", "office_oxide", ["lib/liboffice_oxide.a", "include/office_oxide_c/office_oxide.h"], office_version),]- import tarfile- import zipfile-- def _prune_stale_onnxruntime(static_lib_dir, version):- """Remove ONNX Runtime version dirs under static_lib that do NOT match- `version`. Without this, a version bump leaves the stale dir next to- the new one and build.sh's `find ... -name '*.a'` links BOTH (duplicate- symbols / wrong version, silently)."""- if not os.path.isdir(static_lib_dir):- return- expected = _ort_normalized_dir(version)- for name in os.listdir(static_lib_dir):- if not name.startswith("onnxruntime-linux-x64-static_lib-"):- continue- if name == expected:- continue- stale = os.path.join(static_lib_dir, name)- print(f" Removing stale ONNX Runtime dir: {stale}")- shutil.rmtree(stale)-- for archive, subdir in extractions:++ for archive, subdir, required_files, version in extractions:archive_path = os.path.join(os.getcwd(), archive)if not os.path.isfile(archive_path):print(f" Skipping extraction: {archive} not found")continuetarget = os.path.join(native_deps_dir, subdir)-- # The infiniflow/ragflow-build release zip carries a top-level dir- # named onnxruntime-v{ORT_VERSION}-linux-x86_64. A plain- # "any .a present?" skip would keep a STALE version in place after a- # bump: the new zip downloads, but extraction is skipped because the- # old .a is still under static_lib, so the bump silently does nothing.- # Prune stale version dirs and only skip when the matching version is- # already extracted.- if subdir == os.path.join("onnxruntime", "static_lib"):- _prune_stale_onnxruntime(target, ORT_VERSION)- version_dir = os.path.join(target, _ort_normalized_dir(ORT_VERSION))- if os.path.isdir(version_dir) and any(f.endswith(".a") for _, _, files in os.walk(version_dir) for f in files):- print(f" ✓ {subdir} ({ORT_VERSION}) already extracted to {version_dir}")- continue-- if os.path.isdir(target) and any(f.endswith(".a") for _, _, files in os.walk(target) for f in files):- print(f" ✓ {subdir} already extracted to {target}")- continue- os.makedirs(target, exist_ok=True)- print(f" Extracting {archive} → {target}")- if archive_path.endswith(".zip"):- with zipfile.ZipFile(archive_path) as zf:- zf.extractall(target)- # The infiniflow/ragflow-build release zip carries a top-level dir- # named onnxruntime-v{version}-linux-x86_64, but build.sh's glob and- # the stale checks above all expect- # onnxruntime-linux-x64-static_lib-{version}-glibc2_28. Rename it so- # every consumer shares one name convention (driven by ORT_VERSION).- if subdir == os.path.join("onnxruntime", "static_lib"):- extracted = os.path.join(target, _ort_extracted_dir(ORT_VERSION))- normalized = os.path.join(target, _ort_normalized_dir(ORT_VERSION))- if os.path.isdir(extracted) and extracted != normalized:- if os.path.exists(normalized):- shutil.rmtree(normalized)- print(f" Renaming {os.path.basename(extracted)} → {os.path.basename(normalized)}")- os.rename(extracted, normalized)- else:- with tarfile.open(archive_path) as tf:- tf.extractall(target)+ try:+ _extract_native_archive(archive_path, target, required_files, version)+ except (tarfile.TarError, EOFError, RuntimeError):+ os.unlink(archive_path)+ raise++ _install_stagehand()++ if not extract_onnxruntime(+ os.path.join(native_deps_dir, "onnxruntime", "static_lib"),+ os.path.join(os.getcwd(), _ort_asset_name(ORT_VERSION)),+ ORT_VERSION,+ ):+ # The archive was not downloaded or failed to extract, so no .a landed.+ # Fail loud instead of exiting 0: build.sh's ORT guard would otherwise+ # reject the build later with a less actionable message, and a missing+ # .a left here is exactly the "silent green" this PR is meant to prevent.+ print(+ f" ERROR: ONNX Runtime static archives for {ORT_VERSION} were not "+ f"extracted to {os.path.join(native_deps_dir, 'onnxruntime', 'static_lib')}. "+ f"Check the download above; build.sh will refuse to link without them.",+ file=sys.stderr,+ )+ sys.exit(1)# ONNX Runtime is statically linked into the server binary, so there is no# runtime .so to surface. Log where build.sh (ONNXRUNTIME_STATIC_PREFIX) will@@ -358,38 +610,13 @@ if __name__ == "__main__":if ort_a_files:print(f" ✓ onnxruntime static archives ready: {len(ort_a_files)} .a under {ort_static_dir}")else:- print(f" Skipping onnxruntime static check: no .a found under {ort_static_dir}")-- local_dir = os.path.abspath("nltk_data")- # NLTK >=3.8.2 gates `wordnet` behind `omw-1.4`; both must be provisioned- # or tokenization-backed paths raise LookupError at runtime.- for data in ["omw-1.4", "wordnet", "punkt", "punkt_tab"]:- print(f"Downloading nltk {data}...")- nltk.download(data, download_dir=local_dir)-- for repo_id in repos:- print(f"Downloading huggingface repo {repo_id}...")- download_model(repo_id)-- # Guard: the Go in-process DeepDoc backend loads the .ort weights from the- # InfiniFlow/deepdoc snapshot pulled above. snapshot_download fetches the- # whole repo, so these must be present; fail loudly if a future repo layout- # drops them, so the Go backend can never silently ship without its models.- # (internal/common.DeepDocModelFiles is the authoritative list.)- deepdoc_local = os.path.abspath(os.path.join("huggingface.co", "InfiniFlow", "deepdoc"))- go_model_files = ["det.ort", "layout.ort", "tsr.ort", "rec.ort", "ocr.res"]- if not os.path.isdir(deepdoc_local):- print(- f" ERROR: {deepdoc_local} does not exist; the InfiniFlow/deepdoc snapshot did not materialize.",- file=sys.stderr,- )+ print(f" ERROR: ONNX Runtime .a files still missing under {ort_static_dir} after extraction; build.sh will refuse to link.", file=sys.stderr)sys.exit(1)- missing_models = [f for f in go_model_files if not os.path.isfile(os.path.join(deepdoc_local, f))]- if missing_models:- for f in missing_models:- print(- f" ERROR: expected Go model file {f} missing from {deepdoc_local}; the InfiniFlow/deepdoc snapshot no longer ships .ort weights.",- file=sys.stderr,- )- sys.exit(1)- print(f" ✓ Go .ort model files present under {deepdoc_local}")++ # Download the Go DeepDoc `.ort` weights so this script is a one-stop for Go+ # dev: native libs (above) + model files (below) from a single invocation.+ download_go_models(args.china_mirrors)++ # Same one-stop idea for the token counters: their tokenizer files, in the layout+ # the loaders search and ragflow_deps/Dockerfile ships.+ download_tokenizer_assets(args.china_mirrors)diff --git a/ragflow_deps/download_go_deps.py b/ragflow_deps/download_go_deps.pydeleted file mode 100755index a23b27b0bf..0000000000--- a/ragflow_deps/download_go_deps.py+++ /dev/null@@ -1,535 +0,0 @@-#!/usr/bin/env python3--# PEP 723 metadata-# /// script-# requires-python = ">=3.10"-# dependencies = [-# "huggingface-hub",-# "requests",-# ]-# ///--# This script downloads every artifact the Go build needs: the native static-# libraries (pdfium / pdf_oxide / office_oxide / onnxruntime) for `build.sh`,-# and the Go DeepDoc `.ort` weights (det/layout/tsr/rec.ort + ocr.res) so a Go-# dev can run the in-process backend locally without separately running-# `download_deps.py`. Run it from anywhere — the `__main__` block chdir's into-# this file's own directory, so all outputs land under `ragflow_deps/`-# regardless of the caller's CWD.-#-# Build-context relationship: `ragflow_deps/Dockerfile` is built with-# `ragflow_deps/` as its build context, so the files written here MUST-# sit at the top of `ragflow_deps/`. The Dockerfile's COPY lines assume-# top-level paths (`huggingface.co`, `nltk_data`, `cl100k_base.tiktoken`,-# `*.deb`, `*.jar`, `*.tar.gz`, `stagehand-server-v3-linux-<arch>`).-#-# Typical workflow:-#-# uv run python3 ragflow_deps/download_go_deps.py # download-# cd ragflow_deps-# docker build -f Dockerfile -t infiniflow/ragflow_deps .-#-# The main `Dockerfile` (built from the project root) pulls this image-# via `--mount=type=bind,from=infiniflow/ragflow_deps:latest,...` and-# is unaffected by where these files live locally.-#-# Go DeepDoc weights: in addition to the native libs, this script downloads the-# five Go model files (internal/common.DeepDocModelFiles) from InfiniFlow/deepdoc-# straight into the repo's canonical model directory `internal/rag/res/deepdoc/` (one level-# up from this script). The Go server auto-discovers that directory via-# resolveDeepDocModelDir(), and build.sh --test-native / internal/deepdoc/native/run.sh-# default MODEL_DIR there too — so after running this script NO MODEL_DIR env needs-# to be set:-#-# bash build.sh --test-native # or: cd internal/deepdoc/native && bash run.sh-#-# Embedding tokenizer assets: the Go token counters in internal/tokenizer load a-# tokenizer file per family (XLM-R SentencePiece, BERT WordPiece, two byte-level BPE-# families) from `huggingface.co/<repo>/<file>` at the top of ragflow_deps/ - the path-# the loaders search and the one ragflow_deps/Dockerfile ships. This script fetches-# them too, so one invocation gives a Go checkout everything it counts with. Without-# them every model that declares a tokenizer silently counts with the calibrated-# cl100k estimate instead (see internal/tokenizer/embedding_token_limits.md).--import argparse-import hashlib-import os-import shutil-import sys-import zipfile--import requests--# Mirrors internal/common.DeepDocORTVersion (Go in-process backend). ONE OF-# FOUR places (with that Go constant, ORT_VERSION in ragflow_deps/download_deps.py,-# and ARG ORT_VERSION in Dockerfile_go) that must carry the same ONNX Runtime-# native release for the statically-linked Go DeepDoc backend. There is no-# single source of truth — keep all four equal. build.sh --check-ort-version-# greps this file (and the other three) to fail fast on drift. (The Python pip-# onnxruntime== pin in pyproject.toml is versioned independently and is not-# part of this check.)-#-# Source of the native static archives: infiniflow/ragflow-build (our own-# ORT-only minimal build), NOT the third-party csukuangfj/onnxruntime-libs-# account. The release tag is `onnxruntime-v{ORT_VERSION}` and the asset is-# `onnxruntime-v{ORT_VERSION}-linux-x86_64.zip`. The archive is occasionally-# re-issued under this SAME tag/asset name with patched content; both download-# scripts (this one and download_deps.py) detect that via a `{asset}.sha256`-# sidecar and re-download/re-extract, so a stale local copy never silently-# lingers.-ORT_VERSION = "1.29.0"---def _ort_asset_name(version):- """Release asset filename under infiniflow/ragflow-build tag onnxruntime-v{version}."""- return f"onnxruntime-v{version}-linux-x86_64.zip"---def _ort_extracted_dir(version):- """Top-level directory name INSIDE the release zip (what extractall creates)."""- return f"onnxruntime-v{version}-linux-x86_64"---def _ort_normalized_dir(version):- """Directory name build.sh's `find ... -name '*.a'` glob expects under static_lib."""- return f"onnxruntime-linux-x64-static_lib-{version}-glibc2_28"---def _sha256_of(path):- """sha256 of a file, streamed in chunks so large archives don't blow memory."""- h = hashlib.sha256()- with open(path, "rb") as f:- for chunk in iter(lambda: f.read(1 << 20), b""):- h.update(chunk)- return h.hexdigest()---# Mirrors internal/common.DeepDocModelFiles (Go in-process DeepDoc backend).-# These are the ONLY weights the Go backend loads; the full InfiniFlow/deepdoc-# repo also ships .onnx (Python-only), which this Go-only script deliberately-# skips to keep the download lean.-DEEPDOC_REPO = "InfiniFlow/deepdoc"-DEEPDOC_MODEL_FILES = ["det.ort", "layout.ort", "tsr.ort", "rec.ort", "ocr.res"]--# Embedding tokenizer assets for the Go counters (internal/tokenizer). Each entry is-# (repo, file, kind):-#-# "runtime" - the counters load it in production, so the runtime images must ship it-# (the copy loop in Dockerfile / Dockerfile_base) and the loader must pin-# its SHA-1;-# "oracle" - only scripts/gen_tokenizer_oracle.py needs it, to regenerate the test-# fixtures, so it is deliberately NOT shipped to the image.-#-# ragflow_deps/test_tokenizer_assets.py holds those three places together, because a-# drift here is invisible at runtime: the counter simply becomes unavailable and ingest-# counts with the calibrated fallback.-#-# Fetched per file, not by snapshot: these repos also carry multi-GB weights we do not-# want. A missing asset is NOT fatal (unlike the DeepDoc weights above, without which-# the server cannot run at all).-TOKENIZER_ASSETS = [- # XLM-R SentencePiece (Unigram) - the BAAI bge / multilingual-e5 / m3e /- # gte-multilingual / jina-v3 family shares this vocabulary.- ("BAAI/bge-m3", "sentencepiece.bpe.model", "runtime"),- # BERT WordPiece vocab for the bge-*-en / e5-* / gte-base family.- ("BAAI/bge-large-en-v1.5", "vocab.txt", "runtime"),- # Byte-level BPE families (Qwen3-Embedding, Mistral/Llama embeddings).- ("Qwen/Qwen3-Embedding-0.6B", "tokenizer.json", "runtime"),- ("intfloat/e5-mistral-7b-instruct", "tokenizer.json", "runtime"),- # Cross-check oracles only: the HF tokenizer.json files the same families convert to.- ("BAAI/bge-m3", "tokenizer.json", "oracle"),- ("BAAI/bge-large-en-v1.5", "tokenizer.json", "oracle"),-]---def prune_stale_onnxruntime(static_lib_dir, version):- """Remove ONNX Runtime version dirs under static_lib that do NOT match- `version`. Without this, a version bump leaves the stale dir next to- the new one and build.sh's `find ... -name '*.a'` links BOTH (duplicate- symbols / wrong version, silently)."""- if not os.path.isdir(static_lib_dir):- return- expected = _ort_normalized_dir(version)- for name in os.listdir(static_lib_dir):- if not name.startswith("onnxruntime-linux-x64-static_lib-"):- continue- if name == expected:- continue- stale = os.path.join(static_lib_dir, name)- print(f" Removing stale ONNX Runtime dir: {stale}")- shutil.rmtree(stale)---def has_static_archives(directory):- """True when `directory` holds at least one static archive (.a)."""- return any(f.endswith(".a") for _, _, files in os.walk(directory) for f in files)---def extract_onnxruntime(static_lib_dir, archive_path, version):- """Ensure the ONNX Runtime static archives for `version` sit under- `static_lib_dir`. Returns True when that version is available afterwards- (extracted now or already present), False when the archive is missing.-- The infiniflow/ragflow-build release zip carries a top-level dir named- onnxruntime-v{version}-linux-x86_64, so a present- `static_lib_dir` is NOT evidence that THIS version is extracted: after a- version bump the stale dir is pruned and the new one must be extracted.- """- if not os.path.isfile(archive_path):- print(f" Skipping extraction: {os.path.basename(archive_path)} not found")- return False- prune_stale_onnxruntime(static_lib_dir, version)- version_dir = os.path.join(static_lib_dir, _ort_normalized_dir(version))- if os.path.isdir(version_dir) and has_static_archives(version_dir):- print(f" ✓ onnxruntime/static_lib ({version}) already extracted to {version_dir}")- return True- os.makedirs(static_lib_dir, exist_ok=True)- print(f" Extracting {os.path.basename(archive_path)} → {static_lib_dir}")- with zipfile.ZipFile(archive_path) as zf:- zf.extractall(static_lib_dir)- # The infiniflow/ragflow-build release zip carries a top-level dir named- # onnxruntime-v{version}-linux-x86_64, but build.sh's glob and the stale- # checks above all expect onnxruntime-linux-x64-static_lib-{version}-glibc2_28.- # Rename it so every consumer shares one name convention (driven by- # ORT_VERSION).- extracted = os.path.join(static_lib_dir, _ort_extracted_dir(version))- normalized = os.path.join(static_lib_dir, _ort_normalized_dir(version))- if os.path.isdir(extracted) and extracted != normalized:- if os.path.exists(normalized):- shutil.rmtree(normalized)- print(f" Renaming {os.path.basename(extracted)} → {os.path.basename(normalized)}")- os.rename(extracted, normalized)- return True---def get_urls(use_china_mirrors=False) -> list[str | list[str]]:- if use_china_mirrors:- return [- # stagehand-server-v3 Node.js SEA binaries (used by Browser- # component in local mode).- #- # The stagehand-go Go module (pinned in go.mod) and the- # stagehand-server binary (this release) are LOOSELY- # MATCHED — both stay on the v3.x line and remain- # protocol-compatible. The two version numbers do NOT- # track each other: the Go SDK is at v3.21.0 while the- # current latest server release is v3.7.2.- #- # On every go.mod bump, refresh this URL to the current- # latest server release. There is no version- # correspondence to maintain; "both on v3.x" is the- # compatibility contract.- "https://gh-proxy.com/https://github.com/browserbase/stagehand/releases/download/stagehand-server-v3/v3.7.2/stagehand-server-v3-linux-x64",- "https://gh-proxy.com/https://github.com/browserbase/stagehand/releases/download/stagehand-server-v3/v3.7.2/stagehand-server-v3-linux-arm64",- # Native static libraries for Go build (pdfium, pdf_oxide,- # office_oxide, onnxruntime). Used by build.sh's check_*_deps- # functions — pre-downloaded to avoid network access during CI.- ["https://gh-proxy.com/https://github.com/kognitos/pdfium-static/releases/download/chromium%2F7809/pdfium-linux-x64-static.tgz", "pdfium-linux-x64-static.tgz"],- ["https://gh-proxy.com/https://github.com/yfedoseev/pdf_oxide/releases/download/v0.3.73/pdf_oxide-go-ffi-linux-amd64.tar.gz", "pdf_oxide-go-ffi-linux-amd64.tar.gz"],- ["https://gh-proxy.com/https://github.com/yfedoseev/office_oxide/releases/download/v0.1.12/native-linux-x86_64.tar.gz", "office_oxide-linux-x86_64.tar.gz"],- [- f"https://gh-proxy.com/https://github.com/infiniflow/ragflow-build/releases/download/onnxruntime-v{ORT_VERSION}/{_ort_asset_name(ORT_VERSION)}",- _ort_asset_name(ORT_VERSION),- ],- ]- else:- return [- # stagehand-server-v3 Node.js SEA binaries (used by Browser- # component in local mode).- #- # The stagehand-go Go module (pinned in go.mod) and the- # stagehand-server binary (this release) are LOOSELY- # MATCHED — both stay on the v3.x line and remain- # protocol-compatible. The two version numbers do NOT- # track each other: the Go SDK is at v3.21.0 while the- # current latest server release is v3.7.2.- #- # On every go.mod bump, refresh this URL to the current- # latest server release. There is no version- # correspondence to maintain; "both on v3.x" is the- # compatibility contract.- "https://github.com/browserbase/stagehand/releases/download/stagehand-server-v3/v3.7.2/stagehand-server-v3-linux-x64",- "https://github.com/browserbase/stagehand/releases/download/stagehand-server-v3/v3.7.2/stagehand-server-v3-linux-arm64",- # Native static libraries for Go build (pdfium, pdf_oxide,- # office_oxide, onnxruntime). Used by build.sh's check_*_deps- # functions — pre-downloaded to avoid network access during CI.- ["https://github.com/kognitos/pdfium-static/releases/download/chromium%2F7809/pdfium-linux-x64-static.tgz", "pdfium-linux-x64-static.tgz"],- ["https://github.com/yfedoseev/pdf_oxide/releases/download/v0.3.73/pdf_oxide-go-ffi-linux-amd64.tar.gz", "pdf_oxide-go-ffi-linux-amd64.tar.gz"],- ["https://github.com/yfedoseev/office_oxide/releases/download/v0.1.12/native-linux-x86_64.tar.gz", "office_oxide-linux-x86_64.tar.gz"],- [- f"https://github.com/infiniflow/ragflow-build/releases/download/onnxruntime-v{ORT_VERSION}/{_ort_asset_name(ORT_VERSION)}",- _ort_asset_name(ORT_VERSION),- ],- ]---def download_with_progress(url, filename):- response = requests.get(url, stream=True)- total_size = int(response.headers.get("content-length", 0))- block_size = 1024-- with open(filename, "wb") as file:- downloaded = 0- for data in response.iter_content(block_size):- file.write(data)- downloaded += len(data)-- if total_size > 0:- progress = (downloaded / total_size) * 100- sys.stdout.write(f"\rProgress: {progress:.1f}% ({downloaded}/{total_size} bytes)")- sys.stdout.flush()-- print()---def download_go_models(use_china_mirrors=False):- """Download the Go DeepDoc `.ort` weights so a Go dev can run the in-process- backend with no further setup.-- The files are written into the repo's canonical model directory- `internal/rag/res/deepdoc/` (relative to the repo root, one level up from this- script). The Go server auto-discovers that directory via- resolveDeepDocModelDir(), and build.sh --test-native / run.sh default- MODEL_DIR there too — so after this script runs, no MODEL_DIR env needs to- be set.-- Uses hf_hub_download per-file (not snapshot_download) to fetch only the- five Go model files; the `.onnx` siblings are Python-only and skipped.- On China mirrors, route through hf-mirror.com via HF_ENDPOINT.- """- if use_china_mirrors:- os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"- # Imported lazily so the module stays importable (and its unit tests stay- # huggingface-free) without the huggingface_hub dependency installed.- from huggingface_hub import hf_hub_download-- # Canonical local-dev model dir the Go backend auto-discovers.- repo_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))- target_dir = os.path.join(repo_root, "internal", "rag", "res", "deepdoc")- os.makedirs(target_dir, exist_ok=True)-- missing = []- for fname in DEEPDOC_MODEL_FILES:- dest = os.path.join(target_dir, fname)- if os.path.isfile(dest):- print(f" ✓ {fname} already present")- continue- print(f"Downloading deepdoc model {fname}...")- try:- hf_hub_download(repo_id=DEEPDOC_REPO, filename=fname, local_dir=target_dir)- except Exception as e: # noqa: BLE001 - collected and surfaced below- missing.append((fname, e))-- if missing:- for fname, e in missing:- print(f" ERROR: failed to download {fname}: {e}", file=sys.stderr)- print(- "\n"- "The Go in-process DeepDoc backend loads ONLY the .ort weights listed in\n"- "internal/common.DeepDocModelFiles. They are fetched from "- f"{DEEPDOC_REPO} into:\n"- f" {target_dir}\n"- "Without them the backend cannot serve — the server exits with a fatal\n"- '"no in-process DeepDoc backend serving". To recover:\n'- " - re-run this script (a transient HF/network error usually clears);\n"- " - behind the GFW, re-run with --china-mirrors (routes via hf-mirror.com);\n"- " - or run `uv run python3 ragflow_deps/download_deps.py`, which snapshots\n"- f" all of {DEEPDOC_REPO} (it also provides the Python-side .onnx);\n"- " - or copy the missing files into that directory by hand.",- file=sys.stderr,- )- sys.exit(1)-- print(f" ✓ Go DeepDoc models ready under {target_dir}")- print(" No MODEL_DIR env needed: the Go backend auto-discovers this directory.")---def download_tokenizer_assets(use_china_mirrors=False):- """Download the embedding tokenizer assets the Go counters load.-- Written to `ragflow_deps/huggingface.co/<repo>/<file>`: the path internal/tokenizer- searches and the one ragflow_deps/Dockerfile builds its image from. Fetched per file- (these repos also carry multi-GB weights we do not want), routed via hf-mirror.com- when --china-mirrors is set.-- A missing asset is fatal, like every other asset this script provisions. The counters- would otherwise fall back to the calibrated cl100k estimate, which under-counts some- tokenizers - and an under-count is what makes a provider answer 400. The runtime- images refuse to build on exactly the same condition (the Dockerfile copy loops exit- 1), so provisioning has to fail here rather than hand them a broken tree.- """- if use_china_mirrors:- os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"- # Imported lazily so the module stays importable (and its unit tests stay- # huggingface-free) without the huggingface_hub dependency installed.- from huggingface_hub import hf_hub_download-- base = os.path.join(os.path.dirname(os.path.abspath(__file__)), "huggingface.co")- missing = []- for repo_id, filename, _kind in TOKENIZER_ASSETS:- target_dir = os.path.join(base, repo_id)- if os.path.isfile(os.path.join(target_dir, filename)):- print(f" ✓ {repo_id}/{filename} already present")- continue- os.makedirs(target_dir, exist_ok=True)- print(f"Downloading tokenizer asset {repo_id}/{filename}...")- try:- hf_hub_download(repo_id=repo_id, filename=filename, local_dir=target_dir)- except Exception as e: # noqa: BLE001 - collected and surfaced below- missing.append((repo_id, filename, e))-- if missing:- for repo_id, filename, e in missing:- print(f" ERROR: could not fetch {repo_id}/{filename}: {e}", file=sys.stderr)- print(- "\n"- "The embedding tokenizer counters in internal/tokenizer load these files from\n"- f" {base}\n"- "Without one, the models that declare its counter fall back to the calibrated\n"- "cl100k estimate: safe, but less precise - it under-counts some tokenizers, and\n"- "an under-count is a request the provider rejects. The runtime images refuse to\n"- "build on the same condition (their copy loops exit 1), so this script fails\n"- "instead of handing a broken tree to the image build. To recover:\n"- " - re-run this script (a transient HF/network error usually clears);\n"- " - behind the GFW, re-run with --china-mirrors (routes via hf-mirror.com);\n"- " - or point MODEL_ASSETS_DIR at a directory holding the same layout.\n",- file=sys.stderr,- )- sys.exit(1)-- print(" ✓ embedding tokenizer assets ready")---if __name__ == "__main__":- # Anchor CWD to this file's directory so all relative outputs- # (huggingface.co/, nltk_data/, *.deb, *.jar, *.tar.gz, etc.) land- # at the top of ragflow_deps/ regardless of where the user invokes- # the script from. This is the build context for `ragflow_deps/Dockerfile`.- os.chdir(os.path.dirname(os.path.abspath(__file__)))-- parser = argparse.ArgumentParser(description="Download dependencies with optional China mirror support")- parser.add_argument("--china-mirrors", action="store_true", help="Use China-accessible mirrors for downloads")- args = parser.parse_args()-- urls = get_urls(args.china_mirrors)-- # Some mirrors (e.g. archive.ubuntu.com) reject the default urllib- # User-Agent with HTTP 403, so install an opener with a browser-like UA.- # opener = urllib.request.build_opener()- # opener.addheaders = [("User-Agent", "Mozilla/5.0")]- # urllib.request.install_opener(opener)-- for url in urls:- download_url = url[0] if isinstance(url, list) else url- filename = url[1] if isinstance(url, list) else url.split("/")[-1]- print(f"Downloading {filename} from {download_url}...")-- # The ONNX Runtime archive is re-issued under the SAME release tag and- # asset name whenever its content changes (e.g. the patched build that- # exports SessionGetInitializer*). A pure existence check would then- # keep a colleague's stale local copy and fail to link onnxruntime_go.- # Verify against the published .sha256 sidecar so a re-issued archive- # is always re-downloaded and re-extracted. Every other archive keeps- # the legacy existence-based skip.- is_ort = filename == _ort_asset_name(ORT_VERSION)- expected_sha = None- if is_ort:- sidecar_url = download_url + ".sha256"- try:- resp = requests.get(sidecar_url, timeout=30)- resp.raise_for_status()- expected_sha = resp.text.split()[0]- except Exception as exc: # noqa: BLE001 - best-effort; fall back to legacy- print(f" WARNING: could not fetch {sidecar_url} ({exc}); skipping checksum for {filename}")-- needs_download = True- if os.path.exists(filename):- if expected_sha is not None:- actual = _sha256_of(filename)- if actual == expected_sha:- print(f" ✓ {filename} checksum matches released {expected_sha}; skipping download")- needs_download = False- else:- print(f" {filename} checksum mismatch (local {actual} != released {expected_sha}); re-downloading")- else:- needs_download = False-- if needs_download:- download_with_progress(download_url, filename)- if expected_sha is not None:- actual = _sha256_of(filename)- if actual != expected_sha:- print(f" ERROR: {filename} checksum mismatch after download (got {actual}, expected {expected_sha})", file=sys.stderr)- sys.exit(1)- print(f" ✓ {filename} checksum verified ({actual})")- # Force re-extract below: drop any previously extracted version dir- # so the same-named re-issued archive actually refreshes the .a files.- if is_ort:- native_libs = os.path.expanduser("~/ragflow-native-libs")- version_dir = os.path.join(native_libs, "onnxruntime", "static_lib", _ort_normalized_dir(ORT_VERSION))- if os.path.isdir(version_dir):- print(f" Removing stale extracted ONNX Runtime dir: {version_dir}")- shutil.rmtree(version_dir)-- # Extract native static libraries to ~/ragflow-native-libs for Go build.- # Ensures build.sh can find them without network access.- native_deps_dir = os.path.expanduser("~/ragflow-native-libs")- import tarfile-- extractions = [- ("pdfium-linux-x64-static.tgz", "pdfium-static"),- ("pdf_oxide-go-ffi-linux-amd64.tar.gz", "pdf_oxide"),- ("office_oxide-linux-x86_64.tar.gz", "office_oxide"),- ]-- for archive, subdir in extractions:- archive_path = os.path.join(os.getcwd(), archive)- if not os.path.isfile(archive_path):- print(f" Skipping extraction: {archive} not found")- continue- target = os.path.join(native_deps_dir, subdir)- if os.path.isdir(target):- print(f" ✓ {subdir} already extracted to {target}")- continue- os.makedirs(target, exist_ok=True)- print(f" Extracting {archive} → {target}")- with tarfile.open(archive_path) as tf:- tf.extractall(target)-- if not extract_onnxruntime(- os.path.join(native_deps_dir, "onnxruntime", "static_lib"),- os.path.join(os.getcwd(), _ort_asset_name(ORT_VERSION)),- ORT_VERSION,- ):- # The archive was not downloaded or failed to extract, so no .a landed.- # Fail loud instead of exiting 0: build.sh's ORT guard would otherwise- # reject the build later with a less actionable message, and a missing- # .a left here is exactly the "silent green" this PR is meant to prevent.- print(- f" ERROR: ONNX Runtime static archives for {ORT_VERSION} were not "- f"extracted to {os.path.join(native_deps_dir, 'onnxruntime', 'static_lib')}. "- f"Check the download above; build.sh will refuse to link without them.",- file=sys.stderr,- )- sys.exit(1)-- # ONNX Runtime is statically linked into the server binary, so there is no- # runtime .so to surface. Log where build.sh (ONNXRUNTIME_STATIC_PREFIX) will- # find the archives — the .a files live under- # ~/ragflow-native-libs/onnxruntime/static_lib. The in-process backend- # resolves OrtGetApiBase via dlopen(NULL); there is no dynamic .so fallback.- ort_static_dir = os.path.join(native_deps_dir, "onnxruntime", "static_lib")- ort_a_files = [os.path.join(root, f) for root, _, files in os.walk(ort_static_dir) for f in files if f.endswith(".a")]- if ort_a_files:- print(f" ✓ onnxruntime static archives ready: {len(ort_a_files)} .a under {ort_static_dir}")- else:- print(f" ERROR: ONNX Runtime .a files still missing under {ort_static_dir} after extraction; build.sh will refuse to link.", file=sys.stderr)- sys.exit(1)-- # Download the Go DeepDoc `.ort` weights so this script is a one-stop for Go- # dev: native libs (above) + model files (below) from a single invocation.- download_go_models(args.china_mirrors)-- # Same one-stop idea for the token counters: their tokenizer files, in the layout- # the loaders search and ragflow_deps/Dockerfile ships.- download_tokenizer_assets(args.china_mirrors)diff --git a/ragflow_deps/test_download_deps.py b/ragflow_deps/test_download_deps.pynew file mode 100644index 0000000000..6eb44c0e23--- /dev/null+++ b/ragflow_deps/test_download_deps.py@@ -0,0 +1,253 @@+# Tests for ragflow_deps/download_deps.py ONNX Runtime extraction.+#+# build.sh's build_go() fails fast when libonnxruntime.a is not linked, so these+# tests must guarantee the archive is really landed on disk — above all after an+# ORT version bump, where the zip's version-stamped top-level dir changes but+# static_lib/ still exists from the previous run.+#+# The release zip from infiniflow/ragflow-build carries a top-level dir named+# onnxruntime-v{version}-linux-x86_64; extract_onnxruntime() must rename it to+# the build.sh-expected onnxruntime-linux-x64-static_lib-{version}-glibc2_28 so+# every consumer shares one name convention. These tests pin that rename.++import http.server+import io+import os+import runpy+import shutil+import sys+import tarfile+import threading+import types+import zipfile+from pathlib import Path++import download_deps as deps+import pytest+import requests+from download_deps import (+ _ort_asset_name,+ _ort_extracted_dir,+ _ort_normalized_dir,+ extract_onnxruntime,+ has_static_archives,+)+++def make_ort_zip(path, version):+ """Build a zip shaped like the infiniflow/ragflow-build ORT release: a+ top-level dir named onnxruntime-v{version}-linux-x86_64 holding+ lib/libonnxruntime.a. extract_onnxruntime() must rename it to the+ build.sh-expected onnxruntime-linux-x64-static_lib-{version}-glibc2_28."""+ member = f"{_ort_extracted_dir(version)}/lib/libonnxruntime.a"+ with zipfile.ZipFile(path, "w") as zf:+ zf.writestr(member, b"!<arch>\nort-payload")+ return path+++def test_extracts_on_first_run(tmp_path):+ # static_lib/ does not exist yet: the plain first-run path.+ static_lib = tmp_path / "onnxruntime" / "static_lib"+ version = "1.29.0"+ archive = make_ort_zip(tmp_path / _ort_asset_name(version), version)++ assert extract_onnxruntime(str(static_lib), str(archive), version) is True++ extracted = static_lib / _ort_extracted_dir(version)+ assert not extracted.exists(), "release zip top-level dir must be renamed away"+ version_dir = static_lib / _ort_normalized_dir(version)+ assert version_dir.is_dir()+ assert has_static_archives(str(version_dir))+++def test_extracts_after_version_bump_when_static_lib_exists(tmp_path):+ """Regression: a static_lib/ left over from a previous version must not+ suppress extraction of the bumped version. Pruning the stale dir leaves+ static_lib/ in place; if extraction is then skipped no .a is ever landed and+ build.sh's ORT guard rejects the build."""+ static_lib = tmp_path / "onnxruntime" / "static_lib"+ static_lib.mkdir(parents=True)+ stale = static_lib / _ort_normalized_dir("1.28.0")+ (stale / "lib").mkdir(parents=True)+ (stale / "lib" / "libonnxruntime.a").write_bytes(b"old-ort")++ version = "1.29.0"+ archive = make_ort_zip(tmp_path / _ort_asset_name(version), version)++ assert extract_onnxruntime(str(static_lib), str(archive), version) is True++ assert not stale.exists(), "stale ORT version should be pruned"+ extracted = static_lib / _ort_extracted_dir(version)+ assert not extracted.exists(), "release zip top-level dir must be renamed away"+ version_dir = static_lib / _ort_normalized_dir(version)+ assert version_dir.is_dir(), "bumped ORT version was silently not extracted"+ assert has_static_archives(str(version_dir)), "bumped ORT version landed no .a"+++def test_skips_when_archive_missing(tmp_path):+ static_lib = tmp_path / "onnxruntime" / "static_lib"+ assert extract_onnxruntime(str(static_lib), str(tmp_path / "missing.zip"), "1.29.0") is False+++def test_idempotent_when_matching_version_already_present(tmp_path):+ static_lib = tmp_path / "onnxruntime" / "static_lib"+ version = "1.29.0"+ archive = make_ort_zip(tmp_path / _ort_asset_name(version), version)++ assert extract_onnxruntime(str(static_lib), str(archive), version) is True+ before = sorted(os.listdir(static_lib))+ assert extract_onnxruntime(str(static_lib), str(archive), version) is True+ assert sorted(os.listdir(static_lib)) == before+++def download_server():+ class Handler(http.server.BaseHTTPRequestHandler):+ def do_GET(self):+ if self.path == "/missing":+ self.send_response(404)+ self.end_headers()+ self.wfile.write(b"not an artifact")+ else:+ self.send_response(200)+ self.send_header("Content-Length", "100" if self.path == "/interrupted" else "7")+ self.end_headers()+ self.wfile.write(b"payload")++ def log_message(self, *args):+ pass++ server = http.server.ThreadingHTTPServer(("127.0.0.1", 0), Handler)+ thread = threading.Thread(target=server.serve_forever)+ thread.start()+ try:+ yield f"http://127.0.0.1:{server.server_port}"+ finally:+ server.shutdown()+ thread.join()+ server.server_close()++[email protected]("endpoint", ["missing", "interrupted"])+def test_failed_download_preserves_previous_file(tmp_path, download_server, endpoint):+ dest = tmp_path / "artifact"+ dest.write_bytes(b"previous valid artifact")+ with pytest.raises(requests.RequestException):+ deps.download_with_progress(f"{download_server}/{endpoint}", dest)+ assert dest.read_bytes() == b"previous valid artifact"+ assert list(tmp_path.iterdir()) == [dest]+++def test_successful_download_replaces_previous_file(tmp_path, download_server):+ dest = tmp_path / "artifact"+ dest.write_bytes(b"old")+ deps.download_with_progress(f"{download_server}/complete", dest)+ assert dest.read_bytes() == b"payload"+ assert list(tmp_path.iterdir()) == [dest]+++def test_dependency_preparation_recovers_empty_native_directory(tmp_path, monkeypatch):+ root = Path(__file__).resolve().parents[1]+ context = tmp_path / "ragflow_deps"+ context.mkdir()+ script = context / "download_deps.py"+ shutil.copy(root / "ragflow_deps/download_deps.py", script)+ (tmp_path / "go.mod").write_text("""module fixture++require (+ github.com/yfedoseev/office_oxide/go v0.1.12+ github.com/browserbase/stagehand-go/v3 v3.22.0+)+""")+ native = tmp_path / "native"+ required = {+ "pdfium-linux-x64-static.tgz": ("pdfium-static", ["lib/libpdfium.a", "lib/libc++.a", "lib/libc++abi.a", "include/fpdfview.h"]),+ "pdf_oxide-go-ffi-linux-amd64.tar.gz": ("pdf_oxide", ["lib/linux_amd64/libpdf_oxide.a", "include/pdf_oxide.h"]),+ "office_oxide-linux-x86_64.tar.gz": ("office_oxide", ["lib/liboffice_oxide.a", "include/office_oxide_c/office_oxide.h"]),+ }+ for archive_name, (directory, files) in required.items():+ (native / directory).mkdir(parents=True)+ with tarfile.open(context / archive_name, "w:gz") as archive:+ for name in files:+ item = tarfile.TarInfo("./" + name)+ payload = b"payload\x000.1.12\x00" if name.endswith("liboffice_oxide.a") else b"payload"+ item.size = len(payload)+ archive.addfile(item, io.BytesIO(payload))+ (native / "office_oxide/lib").mkdir()+ (native / "office_oxide/include/office_oxide_c").mkdir(parents=True)+ (native / "office_oxide/lib/liboffice_oxide.a").write_bytes(b"old\x000.1.9\x00")+ (native / "office_oxide/include/office_oxide_c/office_oxide.h").write_bytes(b"old")+ make_ort_zip(context / _ort_asset_name("1.29.0"), "1.29.0")+ for arch in ["x64", "arm64"]:+ (context / f"stagehand-server-v3-linux-{arch}").write_bytes(b"\x7fELFpayload")+ (context / "cl100k_base.tiktoken").write_bytes(b"configured table")++ def offline(*args, **kwargs):+ raise requests.ConnectionError("offline sidecar")++ def hf_download(*, repo_id, filename, local_dir):+ Path(local_dir).mkdir(parents=True, exist_ok=True)+ (Path(local_dir) / filename).write_bytes(b"model")++ monkeypatch.setitem(sys.modules, "huggingface_hub", types.SimpleNamespace(hf_hub_download=hf_download))+ monkeypatch.setattr(requests, "get", offline)+ original_expanduser = os.path.expanduser+ monkeypatch.setattr(os.path, "expanduser", lambda path: str(native) if path == "~/ragflow-native-libs" else original_expanduser(path))+ monkeypatch.setenv("XDG_CACHE_HOME", str(tmp_path / "cache"))+ monkeypatch.setattr(sys, "argv", [str(script)])+ monkeypatch.chdir(tmp_path)+ runpy.run_path(str(script), run_name="__main__")+ for directory, files in required.values():+ for name in files:+ expected = b"payload\x000.1.12\x00" if name.endswith("liboffice_oxide.a") else b"payload"+ assert (native / directory / name).read_bytes() == expected+ stagehand = tmp_path / "cache/stagehand/lib/go_3.22.0/stagehand-server-v3-linux-x64"+ assert stagehand.read_bytes() == b"\x7fELFpayload"+ assert os.access(stagehand, os.X_OK)+ for filename in ["det.ort", "layout.ort", "tsr.ort", "rec.ort", "ocr.res"]:+ assert (context / "huggingface.co/InfiniFlow/deepdoc" / filename).read_bytes() == b"model"+++def test_incomplete_native_archive_preserves_existing_directory(tmp_path):+ target = tmp_path / "native"+ target.mkdir()+ (target / "existing").write_bytes(b"keep")+ archive = tmp_path / "incomplete.tar.gz"+ with tarfile.open(archive, "w:gz"):+ pass+ with pytest.raises(RuntimeError, match="missing required native file"):+ deps._extract_native_archive(archive, target, ["lib/library.a"])+ assert (target / "existing").read_bytes() == b"keep"++[email protected]("kind", ["traversal", "absolute", "symlink", "hardlink", "fifo", "device"])+def test_native_archive_rejects_unsafe_members(tmp_path, kind):+ target = tmp_path / "native"+ target.mkdir()+ (target / "existing").write_bytes(b"keep")+ outside = tmp_path / "outside"+ outside.write_bytes(b"untouched")+ archive_path = tmp_path / "unsafe.tar.gz"+ with tarfile.open(archive_path, "w:gz") as archive:+ library = tarfile.TarInfo("./lib/library.a")+ library.size = len(b"payload")+ archive.addfile(library, io.BytesIO(b"payload"))+ member = tarfile.TarInfo("unexpected")+ if kind == "traversal":+ member.name = "../outside"+ elif kind == "absolute":+ member.name = str(outside)+ else:+ member.type = {"symlink": tarfile.SYMTYPE, "hardlink": tarfile.LNKTYPE, "fifo": tarfile.FIFOTYPE, "device": tarfile.CHRTYPE}[kind]+ member.linkname = str(outside)+ payload = b"unsafe" if member.isfile() else b""+ member.size = len(payload)+ archive.addfile(member, io.BytesIO(payload))++ with pytest.raises(tarfile.TarError, match="Unsafe native archive member"):+ deps._extract_native_archive(archive_path, str(target), ["lib/library.a"])++ assert outside.read_bytes() == b"untouched"+ assert (target / "existing").read_bytes() == b"keep"+ assert not (target / "lib/library.a").exists()+ assert sorted(path.name for path in tmp_path.iterdir()) == ["native", "outside", "unsafe.tar.gz"]diff --git a/ragflow_deps/test_download_go_deps.py b/ragflow_deps/test_download_go_deps.pydeleted file mode 100644index 78c834ad1b..0000000000--- a/ragflow_deps/test_download_go_deps.py+++ /dev/null@@ -1,88 +0,0 @@-# Tests for ragflow_deps/download_go_deps.py ONNX Runtime extraction.-#-# build.sh's build_go() fails fast when libonnxruntime.a is not linked, so these-# tests must guarantee the archive is really landed on disk — above all after an-# ORT version bump, where the zip's version-stamped top-level dir changes but-# static_lib/ still exists from the previous run.-#-# The release zip from infiniflow/ragflow-build carries a top-level dir named-# onnxruntime-v{version}-linux-x86_64; extract_onnxruntime() must rename it to-# the build.sh-expected onnxruntime-linux-x64-static_lib-{version}-glibc2_28 so-# every consumer shares one name convention. These tests pin that rename.--import os-import zipfile--from download_go_deps import (- _ort_asset_name,- _ort_extracted_dir,- _ort_normalized_dir,- extract_onnxruntime,- has_static_archives,-)---def make_ort_zip(path, version):- """Build a zip shaped like the infiniflow/ragflow-build ORT release: a- top-level dir named onnxruntime-v{version}-linux-x86_64 holding- lib/libonnxruntime.a. extract_onnxruntime() must rename it to the- build.sh-expected onnxruntime-linux-x64-static_lib-{version}-glibc2_28."""- member = f"{_ort_extracted_dir(version)}/lib/libonnxruntime.a"- with zipfile.ZipFile(path, "w") as zf:- zf.writestr(member, b"!<arch>\nort-payload")- return path---def test_extracts_on_first_run(tmp_path):- # static_lib/ does not exist yet: the plain first-run path.- static_lib = tmp_path / "onnxruntime" / "static_lib"- version = "1.29.0"- archive = make_ort_zip(tmp_path / _ort_asset_name(version), version)-- assert extract_onnxruntime(str(static_lib), str(archive), version) is True-- extracted = static_lib / _ort_extracted_dir(version)- assert not extracted.exists(), "release zip top-level dir must be renamed away"- version_dir = static_lib / _ort_normalized_dir(version)- assert version_dir.is_dir()- assert has_static_archives(str(version_dir))---def test_extracts_after_version_bump_when_static_lib_exists(tmp_path):- """Regression: a static_lib/ left over from a previous version must not- suppress extraction of the bumped version. Pruning the stale dir leaves- static_lib/ in place; if extraction is then skipped no .a is ever landed and- build.sh's ORT guard rejects the build."""- static_lib = tmp_path / "onnxruntime" / "static_lib"- static_lib.mkdir(parents=True)- stale = static_lib / _ort_normalized_dir("1.28.0")- (stale / "lib").mkdir(parents=True)- (stale / "lib" / "libonnxruntime.a").write_bytes(b"old-ort")-- version = "1.29.0"- archive = make_ort_zip(tmp_path / _ort_asset_name(version), version)-- assert extract_onnxruntime(str(static_lib), str(archive), version) is True-- assert not stale.exists(), "stale ORT version should be pruned"- extracted = static_lib / _ort_extracted_dir(version)- assert not extracted.exists(), "release zip top-level dir must be renamed away"- version_dir = static_lib / _ort_normalized_dir(version)- assert version_dir.is_dir(), "bumped ORT version was silently not extracted"- assert has_static_archives(str(version_dir)), "bumped ORT version landed no .a"---def test_skips_when_archive_missing(tmp_path):- static_lib = tmp_path / "onnxruntime" / "static_lib"- assert extract_onnxruntime(str(static_lib), str(tmp_path / "missing.zip"), "1.29.0") is False---def test_idempotent_when_matching_version_already_present(tmp_path):- static_lib = tmp_path / "onnxruntime" / "static_lib"- version = "1.29.0"- archive = make_ort_zip(tmp_path / _ort_asset_name(version), version)-- assert extract_onnxruntime(str(static_lib), str(archive), version) is True- before = sorted(os.listdir(static_lib))- assert extract_onnxruntime(str(static_lib), str(archive), version) is True- assert sorted(os.listdir(static_lib)) == beforediff --git a/ragflow_deps/test_tokenizer_assets.py b/ragflow_deps/test_tokenizer_assets.pyindex efdf75018d..cc3ed9c214 100644--- a/ragflow_deps/test_tokenizer_assets.py+++ b/ragflow_deps/test_tokenizer_assets.py@@ -3,14 +3,10 @@# The Go counters in internal/tokenizer read their vocabulary from# ragflow_deps/huggingface.co/<repo>/<file>. Three places have to agree:#-# 1. ragflow_deps/download_go_deps.py - fetches the file, and says whether it is-# "runtime" (the counters load it) or "oracle"-# (only scripts/gen_tokenizer_oracle.py does).-# It is the Go-side downloader: everything the-# Go side reads comes from here, and-# ragflow_deps/download_deps.py (upstream) is-# deliberately left alone;-# 2. Dockerfile / Dockerfile_base / Dockerfile_go+# 1. ragflow_deps/download_deps.py - fetches the file, and says whether it is+# "runtime" (the counters load it) or "oracle"+# (only scripts/gen_tokenizer_oracle.py does);+# 2. Dockerfile / Dockerfile_base# - copy the runtime ones into the image at the# path the counters search;# 3. internal/tokenizer/*.go - pin the SHA-1 of the runtime ones.@@ -32,8 +28,7 @@ ROOT = HERE.parentRUNTIME_IMAGES = ("Dockerfile_base", "Dockerfile")GO_SOURCES = ("spm.go", "wordpiece.go", "bpe.go")KINDS = ("runtime", "oracle")-DOWNLOADER = "download_go_deps.py"-UPSTREAM_DOWNLOADER = "download_deps.py"+DOWNLOADER = "download_deps.py"def load_tokenizer_assets():@@ -88,16 +83,3 @@ def test_oracle_only_assets_are_not_shipped():text = image_text(image)for repo, filename, _kind in assets:assert f"{repo}/{filename}" not in text, f"{image} ships {repo}/{filename}, which only the fixture generator needs"---def test_assets_are_fetched_by_the_go_downloader_only():- """The Go-side assets belong to the Go-side downloader.-- ragflow_deps/download_deps.py is an upstream file (it snapshots the Python side's- dependencies); adding the Go counters' tokenizer files there split the list across- two scripts and made the upstream file carry Go concerns. Keeping the fetch in- download_go_deps.py is what makes a Go checkout self-sufficient.- """- upstream = (HERE / UPSTREAM_DOWNLOADER).read_text(encoding="utf-8")- for repo, filename, _kind in load_tokenizer_assets():- assert filename not in upstream, f"{UPSTREAM_DOWNLOADER} mentions {repo}/{filename}; the embedding tokenizer assets are fetched by {DOWNLOADER} only"diff --git a/scripts/gen_tokenizer_oracle.py b/scripts/gen_tokenizer_oracle.pyindex 564551a0c9..9dd793cf66 100755--- a/scripts/gen_tokenizer_oracle.py+++ b/scripts/gen_tokenizer_oracle.py@@ -180,7 +180,7 @@ ORACLES = {# the served behaviour comes from tokenizer.json above.SPM_MODEL = "huggingface.co/BAAI/bge-m3/sentencepiece.bpe.model"-# The cl100k table as shipped in ragflow_deps (the OpenAI blob download_deps.py fetches).+# The cl100k table as provisioned by download_deps.py or build.sh in ragflow_deps.# The Go counter loads this file; tiktoken reads its own cached copy of the same blob, so# the check in main() is what keeps the check-out honest about what it ships.CL100K_TABLE = "cl100k_base.tiktoken"@@ -337,7 +337,7 @@ def main() -> int:else:tokenizer_path = os.path.join(args.deps, rel)if not os.path.isfile(tokenizer_path):- print(f"missing oracle tokenizer: {tokenizer_path}\nrun `uv run ragflow_deps/download_go_deps.py`", file=sys.stderr)+ print(f"missing oracle tokenizer: {tokenizer_path}\nrun `uv run ragflow_deps/download_deps.py`", file=sys.stderr)return 1tokenizer = Tokenizer.from_file(tokenizer_path)vocab = tokenizer.get_vocab()@@ -379,7 +379,7 @@ def main() -> int:# --spm-crosscheck was asked for, so the asset it reads has to be there:# skipping quietly would turn "the second opinion agrees" into a claim# nobody made. Assets are not optional in this script - they are checked.- print(f"missing sentencepiece model: {model_file}\nrun `uv run ragflow_deps/download_go_deps.py`", file=sys.stderr)+ print(f"missing sentencepiece model: {model_file}\nrun `uv run ragflow_deps/download_deps.py`", file=sys.stderr)return 1processor = spm.SentencePieceProcessor(model_file=model_file)divergent = 0