Searching

Search Skill

Do NOT mention internal search strategy or tool selection reasoning to the user. Just search and answer.

Available Search Methods

You have these search methods. Use your judgment to pick the right combination for each query.

search___summarize Hybrid search (vector + keyword) on uploaded documents. Returns matching content with sources.

search___graph_traverse Follows entity connections from search results to discover additional related pages. Requires qa_ids from search___summarize results.

search___graph_keyword Finds pages by keyword similarity in the knowledge graph. Useful when a specific concept or term is the focus.

WebSearch Managed web search (AgentCore). Returns ranked results with source URLs, titles, and publication dates. Use only when the documents and datasets are insufficient. Always cite source URLs.

Structured datasets (Text2SQL)

Some projects also contain structured datasets (spreadsheets converted to queryable tables) alongside unstructured documents. For questions needing exact numbers, aggregation, filtering, ranking, or counting ("how many", "top N", "average", "list all X where ..."), query these instead of (or in addition to) document search.

data___search_datasets Hybrid-search the project's dataset catalog for datasets relevant to a query (returns top matches with name + description + table_name). Call this FIRST with a description of the data you need to discover which dataset(s) can answer the question. Projects may have many datasets, so search rather than list.

data___describe_dataset Return a dataset's reference document (columns, types, enum values, query patterns, caveats) plus its table_name. Call this BEFORE writing SQL — it is the source of truth for column names. Columns are normalized to snake_case (e.g. primary_type, sp_atk), so no double-quoting is needed.

data___run_sql Run a read-only DuckDB query.

• Single dataset: pass dataset_uri; the table is named data (FROM data).
• JOIN across datasets (e.g. multiple sheets, or two related tables): pass dataset_uris (a list) and reference each by its table_name from describe_dataset/search_datasets (e.g. FROM t_aaa a JOIN t_bbb b ON a.code = b.code). Include EVERY dataset your SQL touches in dataset_uris.

Workflow: search_datasets → describe_dataset (for each relevant dataset) → run_sql. Base the answer on query results and show the SQL used.

Joining datasets: when a question needs data combined across tables/sheets, describe_dataset each one to learn its columns and table_name, find the shared key, then issue one run_sql with dataset_uris listing all of them. Prefer a single JOIN query over stitching separate results by hand.

Combinations

• search___summarize alone — quick answer from documents
• search___summarize → search___graph_traverse — deeper search with related pages
• search___graph_keyword alone — explore a concept across documents
• search___summarize + search___graph_keyword — comprehensive search
• search___summarize → search___graph_traverse + search___graph_keyword — maximum coverage
• data___search_datasets → data___describe_dataset → data___run_sql — exact answers from structured data
• Cross question (spans both): document search to identify the entity → data___run_sql with the name/keywords found → synthesize both
• Any of the above + WebSearch — when documents and datasets are not enough

Document/dataset search first. Web search last.

Citations

Use inline citations naturally within the text. For document results, reference the source document and section. For web results, include the URL.

Do not fabricate information or citations that don't exist in search results.