Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

fix: keep OCR text when no image2text model is configured (#18012)

Dhruv Diwakirti86c25068fa37
diff --git a/rag/app/picture.py b/rag/app/picture.py
index 4c6b3cc21c..762d52a719 100644
--- a/rag/app/picture.py
+++ b/rag/app/picture.py
@@ -91,6 +91,7 @@ def chunk(filename, binary, tenant_id, lang, callback=None, **kwargs):
bxs = _get_ocr()(np.array(img))
txt = "\n".join([t[0] for _, t in bxs if t[0]])
+ txt = txt.strip()
callback(0.4, "Finish OCR: (%s ...)" % txt[:12])
if (eng and len(txt.split()) > 32) or len(txt) > 32:
tokenize(doc, txt, eng, language=lang)
@@ -110,6 +111,11 @@ def chunk(filename, binary, tenant_id, lang, callback=None, **kwargs):
tokenize(doc, txt, eng, language=lang)
return attach_media_context([doc], 0, image_ctx)
except Exception as e:
+ if txt:
+ logging.warning(f"CV LLM unavailable, indexing OCR text instead: {e}")
+ callback(msg=f"[WARN] CV LLM unavailable ({e}), indexing OCR text only.")
+ tokenize(doc, txt, eng, language=lang)
+ return attach_media_context([doc], 0, image_ctx)
callback(prog=-1, msg=str(e))
return []
diff --git a/test/unit_test/rag/app/test_picture_video.py b/test/unit_test/rag/app/test_picture_video.py
index 58d6ce5e9f..5923c05ffc 100644
--- a/test/unit_test/rag/app/test_picture_video.py
+++ b/test/unit_test/rag/app/test_picture_video.py
@@ -15,14 +15,22 @@
#
import importlib.util
+import io
import sys
from pathlib import Path
from types import ModuleType, SimpleNamespace
from unittest.mock import patch
+import pytest
+from PIL import Image
-def _load_picture_module(tokenized_texts):
- """Load the picture parser with lightweight fakes for external services."""
+
+def _load_picture_module(tokenized_texts, ocr_text="", vision_error=""):
+ """Load the picture parser with lightweight fakes for external services.
+
+ ``ocr_text`` is what the local OCR returns for an image, and ``vision_error``
+ makes the vision-model lookup fail with that message.
+ """
class FakeLLMBundle:
def __init__(self, *args, **kwargs):
@@ -38,8 +46,15 @@ def _load_picture_module(tokenized_texts):
llm_service = ModuleType("api.db.services.llm_service")
llm_service.LLMBundle = FakeLLMBundle
+ def resolve_default_model(*_args, **_kwargs):
+ if vision_error:
+ # get_tenant_default_model_by_type raises a bare Exception when no
+ # default model is set, not LookupError.
+ raise Exception(vision_error) # noqa: TRY002
+ return {}
+
tenant_model_service = ModuleType("api.db.joint_services.tenant_model_service")
- tenant_model_service.get_tenant_default_model_by_type = lambda *args, **kwargs: {}
+ tenant_model_service.get_tenant_default_model_by_type = resolve_default_model
tenant_model_service.get_first_provider_model_name = lambda *args, **kwargs: None
tenant_model_service.get_composite_model_name_by_id = lambda model_id: model_id
tenant_model_service.resolve_model_config = lambda *args, **kwargs: {}
@@ -54,8 +69,13 @@ def _load_picture_module(tokenized_texts):
string_utils = ModuleType("common.string_utils")
string_utils.clean_markdown_block = lambda value: value
+ def fake_ocr(_image):
+ """Mimic deepdoc's OCR return shape: (box, (text, confidence)) pairs."""
+
+ return [(None, (ocr_text, 0.99))] if ocr_text else []
+
vision = ModuleType("deepdoc.vision")
- vision.OCR = lambda: object()
+ vision.OCR = lambda: fake_ocr
nlp = ModuleType("rag.nlp")
nlp.attach_media_context = lambda docs, *_args: docs
@@ -107,3 +127,97 @@ def test_video_description_is_tokenized_once():
assert len(chunks) == 1
assert chunks[0]["doc_type_kwd"] == "video"
assert tokenized_texts == ["A concise video description."]
+
+
+def _jpeg_bytes():
+ """Encode a small blank image the parser can open."""
+
+ buffer = io.BytesIO()
+ Image.new("RGB", (64, 64), "white").save(buffer, format="JPEG")
+ return buffer.getvalue()
+
+
+def test_short_ocr_text_is_kept_when_vision_model_is_missing():
+ """OCR text under the CV LLM threshold is indexed when no image2text model
+ is configured, instead of being dropped and leaving the image with no chunk.
+
+ Guards https://github.com/infiniflow/ragflow/issues/17941.
+ """
+
+ tokenized_texts = []
+ picture = _load_picture_module(
+ tokenized_texts,
+ ocr_text="工流宛转绕芳甸",
+ vision_error="No default image2text model is set.",
+ )
+
+ callback_calls = []
+ chunks = picture.chunk(
+ "scan.jpg",
+ _jpeg_bytes(),
+ "tenant",
+ "English",
+ callback=lambda *args, **kwargs: callback_calls.append((args, kwargs)),
+ )
+
+ assert len(chunks) == 1, "OCR text was discarded, leaving the image with no chunk"
+ assert chunks[0]["doc_type_kwd"] == "image"
+ assert tokenized_texts == ["工流宛转绕芳甸"]
+
+
[email protected]("ocr_text", ["", " \n"])
+def test_parser_reports_failure_when_ocr_finds_no_text(ocr_text):
+ """With no usable OCR text and no vision model there is nothing to index, so
+ chunk() returns no chunks and reports the failure through the callback."""
+
+ tokenized_texts = []
+ picture = _load_picture_module(
+ tokenized_texts,
+ ocr_text=ocr_text,
+ vision_error="No default image2text model is set.",
+ )
+
+ callback_calls = []
+ chunks = picture.chunk(
+ "blank.jpg",
+ _jpeg_bytes(),
+ "tenant",
+ "English",
+ callback=lambda *args, **kwargs: callback_calls.append((args, kwargs)),
+ )
+
+ assert chunks == []
+ assert not tokenized_texts
+ assert [kwargs.get("msg") for _args, kwargs in callback_calls if kwargs.get("prog") == -1] == ["No default image2text model is set."]
+
+
+def test_ocr_only_fallback_is_reported_as_degraded():
+ """The OCR-only fallback keeps the task successful but must not report a clean
+ parse: the vision failure is surfaced as a [WARN] message instead of prog=-1,
+ so a misconfigured model stays visible to the operator."""
+
+ tokenized_texts = []
+ picture = _load_picture_module(
+ tokenized_texts,
+ ocr_text="invoice total 42",
+ vision_error="No default image2text model is set.",
+ )
+
+ callback_calls = []
+ chunks = picture.chunk(
+ "scan.jpg",
+ _jpeg_bytes(),
+ "tenant",
+ "English",
+ callback=lambda *args, **kwargs: callback_calls.append((args, kwargs)),
+ )
+
+ assert len(chunks) == 1
+ assert [kwargs.get("prog") for _args, kwargs in callback_calls if kwargs.get("prog") == -1] == []
+
+ warnings = [kwargs["msg"] for _args, kwargs in callback_calls if "msg" in kwargs and kwargs["msg"].startswith("[WARN]")]
+ assert len(warnings) == 1, "the vision failure was not reported as a degraded parse"
+ assert "No default image2text model is set." in warnings[0]
+
+ progressed = [args[0] for args, _kwargs in callback_calls if args]
+ assert 0.8 not in progressed, "the fallback still reports a clean 0.8 success"