~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
🔒 Security
- The notification "Test" button's admin-hint decision now shares a single DNS-resolution pass with the default-level URL rejection, closing a DNS-rebinding TOCTOU window that the previous two-call pattern opened between the rejection and the hint check.
✨ New Features
- Added OrcaRouter as a selectable LLM provider. OrcaRouter is an OpenAI-compatible routing gateway; set your
sk-orca-…API key in settings and pick a namespaced model (e.g.orcarouter/autooropenai/gpt-5) to route requests through it.
🐛 Bug Fixes
- A document whose embeddings failed to store is no longer marked as indexed. Previously a chunk-store error left the document flagged indexed but with zero vectors persisted, so it was silently unsearchable and the background reconciler (which only retries un-indexed documents) never picked it up again. The failed index now leaves the document un-indexed so it is retried. (#4855)
- Focused-iteration searches now run inside the request's Flask app context. Previously the per-question searches were dispatched to worker threads with no app context propagated, so any search touching
current_app/graised "Working outside of application context"; the error was swallowed per query and the strategy silently returned no results. It now passescontext_factory=thread_contextlike the source-based strategy. (#4904) - Collection search now ranks results correctly for cosine and dot-product embedding indexes. Previously the relevance score was computed as if every index used L2 distance, which inverted the ranking (the nearest match scored lowest) and could error on an anti-correlated match. (#4963)
- Deleting or cancelling queued research now removes its queue state so it cannot occupy a slot or fail after its parent is gone.
- Settings requests now report server errors instead of treating failed responses as valid settings data.
📝 Other Changes
- Missing inline citations in
langgraph-agentanswers are now diagnosable from the server log: a warning fires when a call skips the citation pass despite accumulated sources (the detailed-report subsection case behind #4969), and when a completed synthesis contains no inline[N]markers despite available sources. Report content is unchanged. (#4969) - Hardened the dynamic npm-update matrix workflow (PR #4997) and added a guardian + behavioral test suite: discovery now prunes node_modules/.claude/.github during traversal, fails loudly on lockfile dirs that are neither the repo root nor under tests/ (instead of silently skipping their build/test), and passes the discovered path through an env var rather than direct ${{ }} interpolation in the change-detection step. Tests pin the wiring (guardian), execute the real find|jq pipeline against synthetic and real-repo trees (behavioral, zero drift), and cover both the empty-matrix and unknown-path fail-loud paths.
- Pytest app fixtures now skip starting the background research queue processor by default, while
LDR_WEB_QUEUE_PROCESSOR_ENABLED=truekeeps an explicit opt-in for tests that exercise queue behavior.