rag-db-advisor
Provides measured advice on using ClickHouse as a vector database backend, including latency, quality comparisons, and operational traps.
Provides measured advice on using Milvus as a vector database backend, including latency, quality comparisons, and operational traps.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
rag-db-advisor

A RAG that answers RAG-stack questions — every claim backed by rag-retriever-bench measurements. Ask it which vector backend fits your workload, what a latency number actually means, or which operational trap you are about to step on. It only answers from measured evidence.
$ rag-db-advisor ask "10万件・日本語・更新頻度高めならどのDB?"Why this exists
Vector-DB comparisons are usually opinions. This one is a closed loop:
Measure — rag-retriever-bench runs 9 backend configurations over the same corpus (MIRACL-ja), same embeddings, same queries, same metrics
Distill — results + the operational traps hit during measurement become the knowledge base (
knowledge/)Serve — this package retrieves that evidence for your question, over MCP or CLI
Dogfood — the retrieval layer imports rag-retriever-bench's own
BaseRetrieverabstraction. The bench data shows every HNSW backend is quality-tied at this corpus size (recall@10 0.979–0.983 @10k), so the store picks the operationally lightest option: Chroma embedded. The advisor follows its own advice.
Related MCP server: mcp-rag-server
Install
pip install git+https://github.com/kenimo49/rag-db-advisor
export OPENAI_API_KEY=sk-... # embeddings only (text-embedding-3-small)
rag-db-advisor ingest # build the local store (54 chunks in v0.1, well under a minute)Use from Claude Code / Claude Desktop (MCP)
claude mcp add rag-db-advisor -- rag-db-advisor mcpTools:
tool | what it does |
| retrieve measured evidence for a free-form question; the calling LLM synthesizes the answer |
| full comparison table at 10k / 100k docs (quality, latency, build time, index verification) |
| operational traps actually hit during measurement |
No generation key needed server-side — MCP returns evidence, your LLM writes the answer.
Use from the CLI
rag-db-advisor ask "ClickHouseのベクトル検索が遅い。何を疑う?" # evidence only
rag-db-advisor ask "pgvectorとQdrantどっち?" --llm # + OpenAI synthesisWhat it knows (v0.1)
9 backend configurations: pgvector, ClickHouse (HNSW ×2 granularities + brute force), Qdrant, Weaviate, Milvus, Chroma, LanceDB
2 corpus scales: 10k / 100k MIRACL-ja passages, 860 human-annotated queries
Operational trap catalog: silent index degradation (3 distinct backends), shared-memory limits, load-visibility issues — each one reproduced, fixed, and written down
Methodology caveats baked into answers: embedded vs server latency is not directly comparable; numbers are MIRACL-ja + text-embedding-3-small on a single node — measure your own data before deciding
Design notes
The advisor never generates verdicts. It returns measured evidence and lets the calling LLM (or human) synthesize the answer — retrieval failures surface as explicit errors so nobody silently falls back to prior knowledge.
Every chunk in the knowledge base traces back to either a bundled bench record (
knowledge/results/*.jsonl) or a hand-written operational note (knowledge/ja/*.md). Notes only cover behavior that was reproduced during the underlying bench work — speculative advice is out of scope by policy.The retriever imports rag-retriever-bench's own
BaseRetrieverand picks Chroma because the bench data says every HNSW backend is quality-tied at this corpus size. The advisor follows its own advice.
Full write-up: docs/methodology.md. Extending the knowledge base: docs/adding-knowledge.md.
License
MIT
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
MCP server for building and testing AI agents with multi-model experimentation and insights.
MCP server providing access to the Scorecard API to evaluate and optimize LLM systems.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceAn MCP server implementation that provides tools for retrieving and processing documentation through vector search, enabling AI assistants to augment their responses with relevant documentation context20 npm263MIT
- AlicenseNot gradedqualityDmaintenanceAn MCP server that indexes documents and serves relevant context to LLMs via Retrieval Augmented Generation (RAG).36MIT
- AlicenseNot gradedqualityBmaintenanceAn MCP server that provides tools for retrieving and processing documentation through vector search, enabling AI assistants to augment their responses with relevant documentation context.10 npmMIT
- AlicenseNot gradedqualityDmaintenanceA benchmarking MCP server that stores facts as both vector RAG and a knowledge graph, enabling head-to-head comparison of memory retrieval methods.MIT