| name | eap-search |
| description | Retrieve exact chunks from the EAP-Runtime offload store. Calls the eap_search MCP tool (lossless, RRF-fused keyword + substring retrieval) and presents the matching chunks with their document id, chunk index, and locator snippet — never a summary in place of the chunk body. Use when the user says "eap search <query>", "search the offloaded output", "find that in the pointer", references a pointer id like eap_ab12…, or runs /eap-search.
|
| license | MIT |
EAP-Runtime search
Call the eap_search MCP tool (eap-runtime server) with the user's query.
How to call
query — the user's terms, as given (the store fuses stemmed-keyword and
substring matching itself; do not pre-mangle the query).queries — optional string array for multi-query fusion in one call.docId — set it when the user names a pointer (eap_…) so retrieval is
scoped to that document.contentType — optional code | prose | all.limit — default 5; raise only if the user asks for more.- Fuzzy correction and proximity rerank are on by default; path-tracked hits
may include
stale: true when the file changed.
Presentation
For each hit show: docId + chunk index (the source span), the snippet as a
locator, and the exact chunk body — retrieval is lossless; never replace
the body with your own summary. Flag stale plainly. You may add interpretation
after quoting. No hits: say so and suggest terms (e.g. from a pointer's
vocabulary) — do not pad with guesses.
Boundaries
- Read-only; never re-index or purge from here.
- If the store is empty or the server unregistered, say so plainly.
