Skip to content

performance.md

Latest commit

 

History

History
63 lines (44 loc) · 2.94 KB

performance.md

File metadata and controls

63 lines (44 loc) · 2.94 KB
Raw

DocIngest Performance Notes

DocIngest has two latency-sensitive paths: the web app for humans and the MCP server for agents. The biggest wins usually come from returning less data, keeping Redis warm, and avoiding filesystem scans on hot routes.

What Is Faster Now

  • GET /api/docs/list returns metadata only, so /view does not download every markdown file just to render the index.
  • GET /api/docs/domain/:domain accepts topic and maxTokens, allowing MCP clients to fetch scoped docs instead of the full markdown blob.
  • The MCP server applies request timeouts, clamps large responses, and caches repeated API calls in-process.
  • The MCP package includes a CLI mode for quick testing without setting up an editor.

MCP Path

The recommended agent flow is:

  1. find-docs to resolve a library name to a domain.
  2. read-docs with topic and maxTokens when the agent needs focused context.
  3. query-docs when the agent needs cross-library examples or snippets.

Useful environment variables:

DOCINGEST_API_URL=https://docingest.com/api
DOCINGEST_TIMEOUT_MS=10000
DOCINGEST_CACHE_TTL_MS=300000

Redis Fast Path

Redis is the recommended self-hosted search/cache layer. The included docker-compose.yml runs Redis locally on host port 6380:

docker compose up -d redis
cd server
npm run build-index

Use REDIS_HOST, REDIS_PORT, and optionally REDIS_PASSWORD to point the backend at local, remote, or production Redis. See Redis setup for the full guide.

If you want the full local crawl/search stack, start the Firecrawl Compose profile instead:

docker compose --profile firecrawl up -d

For every local run mode, see Docker run modes.

High-Impact Next Wins

  • Keep Redis enabled in production and rebuild the search index after new crawls with npm run build-index from server/.
  • Cache GET /api/docs/domain/:domain responses in Redis with ETags, keyed by domain, version, topic, and token limit.
  • Replace filesystem scans on list/search fallback paths with a small generated manifest of domains and metadata.
  • Add a domain-level search endpoint so agents can ask for matching sections inside one docs site instead of reading large files.
  • Lazy-load heavy frontend markdown/rendering dependencies on docs pages, especially syntax highlighting.
  • Virtualize long docs pages and large result lists so the browser only renders visible content.
  • Add CDN or reverse-proxy caching for stable read-only endpoints such as docs metadata, screenshots, and MCP package docs.

Production Checks

  • Confirm /api/docs/fast-search returns source: "redis" for common searches.
  • Confirm /api/docs/list stays metadata-only and does not include raw markdown content.
  • Check browser Web Vitals on /view and /docs/:domain after large corpus imports.
  • Track API response sizes for /api/docs/domain/:domain; large responses usually mean callers should pass topic or maxTokens.