surendra902

Obsidian Vault RAG Assistant — AI skill for Claude Code

AI community

Obsidian Vault RAG Knowledge Assistant: local MiniLM embeddings + numpy retrieval, measured eval harness (recall@5 0.960), Claude or any OpenAI-compatible model, agentic tool loop.

How to install Obsidian Vault RAG Assistant

This entry records only its repository, not the path inside it, so there is no exact command to give. Open surendra902/obsidian-vault-rag-assistant and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Obsidian Vault RAG Assistant does

Obsidian Vault RAG Knowledge Assistant: local MiniLM embeddings + numpy retrieval, measured eval harness (recall@5 0.960), Claude or any OpenAI-compatible model, agentic tool loop.

Alternatives in AI

Claudian — An Obsidian plugin that embeds Claude Code as an AI collaborator in your vault 7.7k ★
Hindsight — State-of-the-art long-term memory for AI agents by Vectorize 6.7k ★
RAG Skill — A Skill dedicated to local knowledge base retrieval 712 ★

README

Obsidian Vault RAG Knowledge Assistant

Ask questions in natural language; get answers grounded in an Obsidian vault, with citations you can open and verify. Local embeddings, a numpy vector store, Claude *or* any OpenAI-compatible model for generation, one tool-use loop for multi-hop questions.

Quickstart (5 commands)

python -m venv .venv && .venv/Scripts/activate     # Windows (source .venv/bin/activate on POSIX)
pip install -r requirements.txt
python index.py --vault ./demo_vault               # ~15s: 27 notes -> 108 chunks
cp .env.example .env                               # then set ONE key; skip for extractive mode
python -m uvicorn app:app                          # http://127.0.0.1:8000

First model load takes ~1 minute (downloads/loads `all-MiniLM-L6-v2`, then cached). Point `--vault` at any Obsidian vault — the indexer is safe against hostile directory shapes (see *Indexer safety* below).

Architecture

vault/*.md
  -> index.py: walk (deny-list) -> dedup (content hash) -> chunk (heading-aware,
                180 words / 30 overlap, token-budget splitter) -> MiniLM embed
                -> vectors.npz + chunks.jsonl + meta.json
  -> app.py (FastAPI):
      POST /ask     cosine top-5 (numpy matmul) -> LLM -> answer + resolved citations
                    top score < 0.27 -> "not in this vault" (no LLM call)
      POST /agent   tool loop (max 8 turns): search_vault / read_note / list_by_tag
      GET  /healthz index stats + provider/model auth probe
      GET  /        chat UI (single static/index.html)

Providers

One key, either kind. Anthropic wins if both are set; with neither the app runs in **extractive** mode (top chunks verbatim, clearly labelled — `/agent` returns 503 because a model has to drive the tools).