~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.
Python academiaanthropicarxivbravedeep-researchencryptionhome-automationhomeserverlocallocal-deep-researchlocal-llmmistralollamaopenaipubmedresearchresearch-toolretrieval-augmented-generationsearxngself-hosted
✨ New Features
- Add an inline, name-based Zotero collection picker with an explicit entire-library option. (#5254)
- Add streaming NDJSON research log export endpoint (
/api/research/<id>/logs/export) and log download action. (#5259)
🐛 Bug Fixes
- Fixed parallel RAG indexing durability and vector index reconciliation. Serialized durable writes to prevent parallel-worker SQLite/FAISS corruption, hardened per-collection FAISS write locks with TOCTOU safety guards, and added refuse-to-shrink guards to index reconciliation to protect existing vector stores during transient failures. (#5235)
- The "Load older" logs re-fetch no longer stamps a false
(2×)duplicate badge on rows that were already rendered; only a genuinely new duplicate occurrence bumps the counter now. (#5253) - Prevent unused OpenAI-compatible endpoint settings from blocking research with other model providers. (#5255)
- Surface benchmark result persistence failures instead of presenting them as an empty successful run. (#5256)
- langgraph-agent
fetch_contenttool: pre-resolve/library/document/<uuid>[/pdf]URLs and bare[N]citation markers before the egress gate so library-only research runs can actually read their sources. The library RAG and collection search engines emit/library/document/<uuid>as a result's citation URL, and the LLM sometimes pastes the[N]citation marker back into the tool; the egress policy correctly rejected both asunsupported_scheme, so every "Reading the page" milestone on a library-only run produced no content. The pre-resolution passes library URLs through a local DB read ofDocument.text_contentand rewrites citation markers viaSearchResultsCollector.find_by_index. (#5266) - Fixed database lock contention in SQLite queue processor v2 by batching updates into single transactions and deduplicating per-research updates. (#5268)
- Reduce false-positive WARNING/ERROR log noise during normal research runs: filter werkzeug Socket.IO disconnect
write() before start_responseerrors, demote DownloadService init path logs to debug, skip the deprecated display-label fallback warning for plain engine keys likecollection_<uuid>, and include exception context in document-URL warnings. (#5272) - langgraph-agent
fetch_contentno longer mis-renders egress denials as📄 From the page: Cannot fetch …in the chat panel — denial and error strings from the tool are now logged as warnings only, and the MILESTONE channel is reserved for successful reads. Thepolicy denied URL fetchWARNING emitted bysecurity/egress/policy.pynow inlines the offending URL (redacted toscheme://host:port) into its message body — alongsidereason,scope,count, andcounted— so a triager reading the persistedresearch_logs_<uuid>.jsonlcan correlate each denial with the URL the agent attempted. With the chat-panel MILESTONE suppression, this WARNING is now the only persistent record of a denial — if the URL were missing here, there would be nothing else to read it from. The message body is whatdatabase_sinkwrites toapp_logs.message, so this is the only layer at which the URL actually reaches the JSONL. (#5278) - Detailed-mode runs no longer log a misleading
format_document_split appears to have over-strippedwarning when the answer body is legitimately smaller than the appended sources list (e.g. runs with hundreds of sources);IntegratedReportGeneratornow wraps the appended## Sourcesblock in a sentinel that the splitter keys off, andreport_contentis correctly persisted as the answer body only. The sentinel is matched at its last occurrence so a spoofed earlier occurrence inside LLM-generated prose cannot silently truncate the answer or disable the over-strip safety check, and the quick-mode save path no longer trusts the sentinel (only the detailed-mode report generator emits it). (#5283) - Replaced manual
database is lockedretry loop withtenacityin_update_task_status. The five-attemptfor attempt in range(5)withtime.sleep(0.05 * (2**attempt))was rewritten usingtenacity.retry(stop=stop_after_attempt(5), wait=wait_exponential(multiplier=0.05), retry=retry_if_exception(_is_database_locked_error)), matching the rest of the codebase's retry idiom. The predicate limits retries to SQLitedatabase is lockederrors; any other exception still escapes on the first attempt so the outer caller can log it without further retry. Total wall-clock backoff is unchanged (~0.05+0.10+0.20+0.40 s before the final attempt). Pins: three regression tests intests/research_library/routes/test_rag_routes_background_coverage.pyverify retry-then-success, fail-fast on non-lock errors, and exhaustion logging. - Stop
safe_rollbackfrom masking the original failure with a secondaryInvalidRequestErrororInterfaceErrorwhen the thread-local SQLAlchemy session is in a broken state — either mid-provisioning on the per-userQueuePool(concurrent-operation race) or holding a cursor invalidated by acommit/rollbackthat fired betweenexecute()and the lazy-attribute fetch. Both states now reset the thread-local cache (so the next caller on the same thread gets a fresh session from theQueuePool) and log at DEBUG; otherSQLAlchemyErrorfailures still hit the loudlogger.exceptionpath (unless their message matches a known broken-session signature). - Persist a short
[Type: value]exception prefix toapp_logsrows fromdatabase_sinkwhen the loguru record carries an active exception, so an investigator queryingResearchLogcan distinguish ERROR lines that share a bare message. The full traceback remains off the encrypted-DB sink per the #4182 password-redaction policy. - Standardize fallback and exception handlers in
TokenCountingCallback._save_to_dbandSearchTracker.record_searchto uselogger.warningwithscrub_errorsanitization, ensuring DB/metrics failure logs are captured safely without exposing tracebacks, credentials, or sensitive encryption passwords on stderr.
- Stop
- LangGraph's primary search tool now identifies the selected source classification to agents.
📝 Other Changes
- Zotero connection, picker, and sync actions now report progress and results through the app-wide notification banner. (#4961)