corpus-search

Corpus Search

Purpose

Run quick passage-level search across one or more ike corpora and return citation-linked results. Optimised for the "just find me the relevant bits" case — not for sustained analysis.

For deeper multi-source analysis, hand off to the academic-review skill. For claim evaluation with a verdict scale, hand off to source-verification. To navigate a single document's structure, use document-navigator.

Before Starting

Call get_server_instructions once per session. The server may return corpus-specific search guidance (scope constraints, house rules on query phrasing) that overrides the defaults below. Apply that guidance for the rest of the session.

Process

1. Discover corpora if scope is unclear. Call list_corpora to see available collections when the user hasn't named one. If the user's request is effectively "show me what I have", display the corpora list with document counts and stop — ask what they'd like to search for before running a query.

2. Identify corpus scope. If the user named a corpus (exact name, slug, or partial match), match it against the corpus list case-insensitively and record the corpus ID. Multiple ambiguous matches: show the candidates and ask the user to pick one. No name mentioned: search all accessible corpora.

3. Run the search. Call search_documents with the user's query text, top_k: 10, and the optional corpus_ids filter. Start with the user's phrasing verbatim; only reformulate if the first pass returns nothing useful.

4. Gather citation links. For each returned result, record the title, short_id, page range, a 1–2 sentence excerpt of the matching passage, and the relevance score. Every search result includes a citation_text and citation_link — use those directly. Call get_citation_link explicitly only when a needed link isn't already in the result payload.

5. Offer next steps. After presenting results, offer:

   • Fetch full content of a specific document (get_document_content)
   • Read a particular section in detail (get_document_section)
   • Get full metadata for any result (get_document_metadata)
   • Run a refined search with narrower or alternative terms

Output Format

Present results as a numbered list. Each entry should be independently readable:

## Search Results: "<query>"
*Searched N corpora, M documents*

1. **Document Title** (`short_id`) — relevance: high/medium/low
   Pages N–M
   > Matching excerpt text here...
   [View document](citation-link)

2. ...

---
Want me to read any of these in detail, or refine the search?

Use citation_text as the visible label inside each markdown link whenever the server provided one — the MCP get_server_instructions rules require human-readable citation labels, not bare URLs.

Edge Cases

• No results: Do not silently report zero hits. Suggest two or three alternative phrasings (synonyms, broader terms, removing a corpus filter) and offer to re-run. If the user clearly scoped to a single corpus, also offer to broaden to all corpora.
• Ambiguous corpus name: Show the matching candidates with IDs and ask the user to pick one before searching.
• Very broad query: Still return results, but flag that the query was broad and suggest specific narrowing terms (names, dates, technical vocabulary). Offer to re-run.
• Tool error on list_corpora or search_documents: The MCP server may be unreachable, misconfigured, or returning auth errors. In that case, recommend the connection-test skill rather than retrying blindly.

Related Skills

• academic-review — for structured multi-source analysis when search results reveal a substantial research question.
• source-verification — to evaluate whether a specific claim is supported by what surfaces in search.
• document-navigator — to interactively explore a single document after a promising hit.
• connection-test — if list_corpora or search_documents itself returns a tool error.