rag-expert

name	rag-expert
description	Local vector search via LanceDB + sentence-transformers, shared as a service for the whole fleet

RAG Expert — local-llm-mcp

This server hosts an optional RAG-as-a-service endpoint that lightweight fleet MCP repos can use instead of shipping their own vector stack. The service uses all-MiniLM-L6-v2 embeddings stored in LanceDB.

When to use

Situation	Use shared RAG?
Repo has < 500 lines of Python	Yes — skip lanceDB + sentence-transformers dep
Repo needs cross-corpus search	Yes — one index, one query
Repo has private/sensitive data	No — keep local RAG
Repo has custom chunking/metadata	No — keep local RAG
Repo already has LanceDB	No — don't change

API

Method	Path	Body / Params	Returns
POST	/api/v1/rag/ingest	{"text", "source"?, "metadata"?}	{"success", "total_docs"}
GET	/api/v1/rag/search	?query=&limit=	[{text, source, metadata, score}]
DELETE	/api/v1/rag/clear	—	{"success", "cleared": true}
GET	/api/v1/rag/status	—	{"available", "model", "total_docs"}

Examples

import httpx
rag = "http://localhost:10833/api/v1/rag"

# Ingest
httpx.post(f"{rag}/ingest", json={"text": "Attention is all you need...", "source": "arxiv:1706.03762"})

# Search
r = httpx.get(f"{rag}/search", params={"query": "transformer paper", "limit": 3})
for hit in r.json():
    print(f"  [{hit['score']}] {hit['source']}: {hit['text'][:80]}...")

Recipes

Ingest from a URL (fetch + index)

import httpx

def ingest_url(url: str):
    resp = httpx.get(url, timeout=30)
    resp.raise_for_status()
    text = resp.text[:50000]  # cap at 50k chars
    httpx.post("http://localhost:10833/api/v1/rag/ingest", json={
        "text": text,
        "source": url,
    })

Search and return matched sources

def search_sources(query: str) -> list[str]:
    r = httpx.get("http://localhost:10833/api/v1/rag/search", params={"query": query, "limit": 10})
    return list(set(h["source"] for h in r.json() if h.get("source")))