- Home
- Skills
- Technical Documentation
- docs-rag: Semantic Search Over Your Docs
Works with the AI Agents you already use
docs-rag: Semantic Search Over Your Docs
Local semantic search over your markdown docs and ADRs that never returns a superseded decision as current truth.
Free
docs-rag: Semantic Search Over Your Docs
Example session with this skill installed
We're about to touch the auth middleware. How long should user sessions last? Set the session cookie maxAge accordingly.
- Read your context and instructions
- Compiled the docs-rag
Answer: sessions last 7 days, no sliding renewal → maxAge: 604800
I checked the project docs with search_notes before touching the code.
What the search returned
[both 0.788] decisions/2026-10-session-length.md
› How long user sessions last [2026-10-12]
[exact match] reference/auth.md
› How the client refreshes the access token
[exact match] decisions/2026-10-session-length.md
› Why we shortened sessions to 7 days
The top hit is labelled [both]: it matched by meaning and on exact terms, the strongest signal. It's the current decision from October 2026:
User sessions last 7 days, with no sliding renewal. After 7 days the user must log in again, regardless of activity. The session cookie maxAge is 604800 seconds.
Why not 30 days?
Your docs also contain an older decision that says "30 days with sliding renewal". It scores almost the same (0.785 vs 0.788), so a plain vector search would have handed it to me just as confidently. It's marked status: superseded, so it was excluded by default. Re-running with searchAll shows it only as history:
[both 0.785] decisions/2025-11-session-length.md
› How long user sessions last
[superseded → decisions/2026-10-session-length.md]
It points to the doc I used, so the chain checks out.
The change
reference/auth.md says cookie options live in server/auth/session.ts:
| Option | Before | After |
|---|---|---|
maxAge | 2592000 (30 d) | 604800 (7 d) |
rolling | true | false (no sliding renewal) |
Next steps
- Existing sessions issued with 30 days stay valid until they expire. Tell me if you want to revoke them now.
- If the policy changes again, tell me and I'll record the new decision, mark this one superseded and re-index.
Connects securely to your tools. The creator never sees your data.
About this skill
Your agent searches your markdown specs, ADRs and decision logs by meaning, fully on your machine, and stops treating dead decisions as current.
This skill pairs with the open-source MCP server docs-rag-mcp (MIT, on npm) and teaches the agent to use it properly:
- Setup, step by step. Ollama +
bge-m3, per-project scaffold, index, verify. No guessed env vars or config keys. - Superseded decisions stay dead. Docs marked
status: supersededdrop out of default search. When the agent hits one, it followssuperseded_byto the current doc instead of quoting history as fact. - "I don't know" beats a confident guess. Below the score threshold the tool answers No relevant results. The skill makes the agent treat that as an answer: rephrase once, check the code or ask, never invent.
- Results it can weigh. Every hit says why it's there:
[semantic 0.71],[both]or[exact match]. - Decisions recorded properly. When a decision changes, the agent writes the new doc, marks the old one superseded, repoints older references and re-indexes.
- Docs written to be found. Includes the authoring guide: how heading structure decides what retrieval can find.
Requirements
- A local coding agent with MCP and a terminal: Claude Code, Cursor, GitHub Copilot / VS Code, Codex CLI, Gemini CLI, Windsurf, Cline…
- Node.js ≥ 22.5 and Ollama running locally.
- Not for browser-only chat apps (ChatGPT web, Claude.ai web): they can't run a local MCP server.
Privacy
Embeddings run through your local Ollama. Nothing leaves your machine at index or query time.
When you don't need it
For exact symbols in code, grep is faster. With a handful of docs, just read them. This is for long, layered docs where outdated decisions are mixed in with current ones.
Does not index code symbols; use grep for that. Requires specific YAML frontmatter. Manual re-indexing needed after edits unless using specific hooks.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
1 use
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 29 days ago
- Free with an Agensi account
Requires
Permissions
Allowed Hosts
File Scopes
Shell: installs the docs-rag-mcp npm package (pinned 0.1.0) and runs its index/search/scaffold commands. Network: npm install only, plus Ollama on localhost:11434 for embeddings. Nothing is sent to external services. Write: per-project config (.rag/, .mcp.json, .gitignore, optional Claude Code re-index hook in .claude/settings.local.json) and lifecycle frontmatter (status, superseded_by) in markdown docs when a decision changes.