Knowledge base MCP server · native RAG hybrid retrieval
Spin up one knowledge base every agent shares — no one rebuilds from scratch.
A single binary, one command, MCP over HTTP built in: Claude, Cursor, or your own agent just connects to one endpoint and shares the same knowledge base. Chinese-native hybrid retrieval, end-to-end traceability, and workspace isolation are what make it dependable.
$ make standalone ✔ langhuan started (single binary, zero deps) ✔ SQLite db + keys auto-provisioned ✔ MCP over HTTP at :8080/mcp $ curl :8080/mcp -d '{"tool":"knowledge_search","query":"how to configure hybrid retrieval"}' { "hits": [{ "content": "vector + full-text two-path recall…" }] }
Why · Why Langhuan exists
These are the common ones, and how Langhuan answers each.
When the second and third agents ship, ingestion, permissions, and retrieval are rebuilt each time, producing inconsistent sets of facts.
Langhuan — Distill knowledge processing into one layer of infrastructure, served over MCP to every agent, reusing the same facts.
How this gets solvedPure vector retrieval can’t reliably find contract codes, product names, or proper nouns — keywords get swallowed by the whole sentence.
Langhuan — zhparser / gse Chinese tokenization plus two-path vector recall, fused with deterministic RRF — exact terms and semantics both hit.
How hybrid search worksWhich document, which version did this answer come from? You can’t say, so debugging is guesswork and auditing is impossible.
Langhuan — Every chunk carries a source anchor (document / version / page); every retrieval has a stable search_id that admins can replay by original query.
You want to validate value first, but PostgreSQL + Redis + multiple containers stop you at the door.
Langhuan — One binary plus one .db file, zero-config — no Docker, PG, or Redis. Download and run.
DownloadThe knowledge base has nothing, but the model confabulates confidently — the classic RAG hallucination.
Langhuan — Langhuan returns evidence, never fabricates answers: retrieval_status distinguishes available / empty / degraded, so you know whether nothing was recalled or there truly is no answer.
Position · Where Langhuan sits
When your agents grow from one to many, the knowledge base shouldn’t be rebuilt each time. Langhuan spins up one knowledge base that every agent connects to over MCP; Chinese-native hybrid retrieval, end-to-end traceability, and single-binary zero-config deployment are what make it dependable. It generates no answers and orchestrates no agents — that is the essential difference from app platforms like Dify/RAGFlow: it sits beneath them.
Highlights
No LLM answer generation, no chat orchestration — just taking "document → traceable retrieval" as far as it can go, as the knowledge foundation for any LLM app, MCP client, or agent.
Chinese-native hybrid retrieval
pgvector embeddings + PostgreSQL FTS (zhparser), or SQLite + gse Chinese tokenization, with two-path recall and deterministic RRF fusion. Chinese keywords are no longer treated as a whole sentence; recall quality works out of the box.
Websites as content sources
Three ingestion modes (crawl/sitemap/urls), four engines (http/render/firecrawl/jina with auto escalation), pages indexed alongside uploaded files; three-tier change detection, robots compliance, global in-flight isolation.
First-class MCP over HTTP
Tools like knowledge_search and document_ingest are exposed directly to MCP clients. Claude, Cursor, and others can use them immediately, with no extra bridge.
Single binary, full stack
REST + MCP + async workers + Web Console embedded in one binary (go:embed). Zero-config: one binary plus one .db file, no PostgreSQL, no Redis.
Fully traceable
Every chunk traces back to its source document, version, and page/line/offset anchor. Ingest, chunking, and indexing are idempotent end to end, with the database as the source of truth.
Workspace-scoped isolation
A workspace is the tenant boundary. Member roles plus API keys scoped to specific knowledge bases provide fine-grained auth, and unauthorized access consistently returns 404.
A library, retrievable at will
Document → normalized fact layer → searchable projection, atomically published to a single active Generation. The knowledge base is the library; retrieval is fetching a book.
Architecture
Baseline:Go 1.26 · Gin · GORM · PostgreSQL 17 + pgvector · FTS · asynq + Redis · SQLite standalone (sqlite-vec + FTS5 + gse)
Quick Start
No Docker, PostgreSQL, or Redis: run the langhuan binary directly and it auto-provisions a SQLite database, keys, and config on first start. Open http://127.0.0.1:8080 to initialize, then ingest and search. For production concurrency, use the PostgreSQL + Redis deployment.
# Zero external dependencies; no Docker / PostgreSQL / Redis $ git clone https://github.com/amoydavid/langhuan.git $ cd langhuan && make standalone
Create an admin account
Create a knowledge base, upload pdf/md/txt/csv/xlsx/docx, or attach web sources that crawl pages automatically
Consistent retrieval across REST / Web / MCP
Positioning
Langhuan pushes "knowledge processing and retrieval" as far as it can go, then exposes it to upper-layer apps over MCP.
| Langhuan | Dify / RAGFlow and similar platforms | |
|---|---|---|
| Position | Knowledge layer (no LLM orchestration) | App platform (with Chat / Agent) |
| Keyword retrieval | Native zhparser / gse tokenization + RRF | Vector-dependent, weaker FTS |
| Delivery | Single binary + one .db file (zero-dep option) | Multi-container suite |
| MCP | Native MCP over HTTP | Extra bridge required |
| Traceability | chunk → page/line anchor, full chain | Partial |
| Integration | As the knowledge foundation for your app | In-platform closed loop |
Integrations
FAQ
Langhuan is not an app platform — it is a knowledge layer. It does not generate LLM answers or orchestrate chat/agents; it takes "document → traceable retrieval" to production grade and exposes it to upper layers over MCP.
Yes, and natively. The full-text path uses zhparser (production) or gse (standalone) Chinese tokenization, and after two-path recall with embeddings, results are fused with deterministic RRF.
Langhuan serves MCP over HTTP natively at /mcp, using a workspace API key as a Bearer token. MCP clients like Claude and Cursor just need one endpoint to call it.
Zero-config: run the langhuan binary directly and it auto-provisions a SQLite database and keys — no Docker, PostgreSQL, or Redis. For production concurrency, use PostgreSQL + Redis.
Download a prebuilt binary for macOS (Apple Silicon / Intel), Linux, or Windows — unzip and run, no Docker, PostgreSQL, or Redis required. See the download page on this site.
Journal · Blog
An everyday question like "埃及有哪些民族？" returns literally nothing from SQLite FTS5 Chinese full-text search — and hybrid retrieval quietly degrades to vector-only, with no errors, no alerts, and normal-looking metrics. The mechanism, the query-side fix, and how to check whether your FTS channel is already spinning in neutral.
The metric that matters most for a knowledge base product is retrieval quality, and all we had were functional tests. How Langhuan built a retrieval eval with 200 human-annotated real queries, a dual-track channel matrix, and bit-identical reproducibility — plus the production bugs it caught on day one and the hybrid-search hypothesis it finally confirmed with data.
A local RAG knowledge base does not need a full database and queue stack on day one. Start with real documents, Chinese-aware retrieval, and evidence; move to production components when the workload earns them.
One command to start, every agent plugged in. Chinese-native hybrid retrieval, end-to-end traceability, single-binary zero-config.