Context7 Documentation Lookup

Answer from fetched docs. If Context7 returns nothing useful, say so rather than silently falling back to memory. Every query is sent to an external API, so it never carries the user's pasted code, logs, internal URLs, secrets, or personal data.

Phase 1 — Resolve the library ID

Call context7:resolve-library-id with:

• libraryName: the library name extracted from the user's question
• query: what to look up in the docs, rewritten in your own words (improves relevance ranking), e.g. Next.js middleware redirect based on auth cookie, not the user's message with its pasted code

Phase 2 — Select the best match

From the resolution results, choose per the tool's own selection guidance. If the user mentioned a version (e.g. "React 19"), prefer version-specific IDs.

Phase 3 — Fetch the documentation

Call context7:query-docs with:

• libraryId: the selected Context7 library ID (e.g. /vercel/next.js)
• query: one topic per call, rewritten the same way as in Phase 1. When the question spans several topics, make one call per topic (at most 3 per question), unless it is about how the topics interact.
• language: the language the user writes in (e.g. TypeScript), when known

Phase 4 — Draft the answer

Draft the answer from the fetched docs: include their code examples and cite the resolved version.

Phase 5 — Verify

Before sending, confirm the answer rests on the fetched docs, not on memory: every API name or option mentioned appears in the query-docs result, and the version cited is the one that was resolved. Re-query or remove any claim the docs don't support, then recheck. If resolution or the query returned nothing useful, say so explicitly rather than answering from training data as if it were current.