Changelog
Tracking how Langhuan evolves across knowledge bases, MCP over HTTP, RAG retrieval, PDF parsing, and the Web Console: why each release shipped, what it solves, and how it makes the knowledge infrastructure more dependable.
Introduces in-KB "web content sources" — three ingestion modes (crawl/sitemap/urls), four fetch engines (http/render/firecrawl/jina with auto escalation), pages indexed alongside uploaded files and retrievable in the same search; each source is independently configurable, re-runnable and deletable, backed by full-stack E2E and real-provider verification.
Four open retrieval directions closed by controlled experiments (semantic chunking, LLM context headers, chunk parameters, traditional/simplified normalization), then progressive disclosure for search results (detail full/lean) and discovery tools for agents (MCP tools 7→10); matched_children no longer carries child content.
Fixes three defects — standalone knowledge base creation returning 500, vector search unavailable in standalone, and FTS zero recall on Chinese questions — plus a disable-per-channel diagnostic semantic and an offline retrieval eval harness.
Without Redis, tasks now queue in SQLite or PostgreSQL — restarting no longer drops jobs; also fixes a standalone OIDC panic and trims the binary ~22%.
Run as a single binary with a single .db file, no PostgreSQL or Redis; coexists with the production PG path, with complete retrieval and zero PG regressions.
Every retrieval now has a stable identity, verifiable citations, and full lineage. Admins can replay a search with the original query within the retention window — no more guessing at RAG recall drift.
Failure retry, Generation rebuild, OpenTelemetry observability, and asynq task governance ship — giving Langhuan the long-term runnability and recoverability needed to enter release readiness.
Stable incremental document sync, safe increment cursors, forced re-sync, and deletion policies ship, so large collaboration spaces sync without dropping content or deleting the wrong thing.
OIDC login, JIT user creation, external-identity binding, and single-tenant auto-join ship, rounding out the enterprise onboarding path.
Sidebar hierarchy and page backgrounds are tightened and dev-only floats removed, so daily management stays focused.
Feishu app management, source sync, and a file-tree browser ship, bringing team docs into Langhuan's retrieval pipeline.
A built-in OpenAPI live doc and Scalar browser ship, so the REST management surface can be explored and verified directly.
A model catalog, multi-capability providers, and a rerank loop ship, so results after hybrid recall land closer to the real question.
Adaptive parent-child chunking and parent-block aggregation ship, so retrieval hits smaller fragments while still returning full context.
Introduces MinerU Cloud PDF parsing, S3-compatible object storage, and document asset archiving, so complex documents keep a traceable structure.
Langhuan ships under the MIT License, bringing REST, MCP over HTTP, and an embedded console together in one runnable binary.