Search rulebook clauses by similarity
search_rulebook_vectorRead-onlyIdempotent
Search rulebook clauses using natural language queries when exact IDs are unknown. Returns ranked matches with similarity scores for citing.
Instructions
Search the curated rulebook clauses by lexical-vector similarity.
Use this when you know *what* a rule is about but not its exact
``scheme``/``version``/``clause`` id: describe it in natural language and
get back the closest curated clauses, each with a similarity ``score``.
Then pass the winning ``scheme``/``version``/``clause`` to
``cite_rulebook`` for the full citation, or browse everything with
``list_rulebook_clauses``.
Retrieval is a **deterministic lexical-vector cosine** search over the
same curated SEPA / CBPR+ / HVPS+ summaries that back ``cite_rulebook``
(no external, copyrighted, or auth-gated rulebook text is indexed). Each
clause and the query are hashed into a fixed 256-dimension term-frequency
vector (whole words plus character 3/4-grams, BLAKE2b-bucketed so results
are reproducible across processes) and ranked by cosine distance with
``sqlite-vec``. It is offline and does **not** use a large neural
embedding model, so the same query always yields the same ranking and no
model download or network call happens at query time.
``sqlite-vec`` ships in the optional ``[vector]`` extra and is imported
lazily; when it is not installed this returns a graceful
``{"error": ...}`` payload asking the operator to
``pip install 'camt053-mcp[vector]'`` rather than failing to import.
Args:
query: The natural-language search string.
top_k: The maximum number of clauses to return (default ``5``,
clamped to the corpus size).
Returns:
``{"query", "top_k", "returned", "method", "results", "disclaimer"}``
where ``results`` is the ranked list of clause dicts (each with an
added ``score``), or an ``{"error": ...}`` payload on a bad argument
or a missing ``[vector]`` extra.
Input Schema
TableJSON Schema
| Name | Required | Description | Default |
|---|---|---|---|
| query | Yes | A natural-language search string (e.g. 'structured address requirement' or 'instant payment settlement time'). Matched against the curated SEPA / CBPR+ / HVPS+ clause summaries. | |
| top_k | No | Maximum number of clauses to return, ranked most-similar first. Clamped to the corpus size; must be positive. Defaults to 5. |