Wiki Agent Search

wiki_agent_search is a dynamic QueryEngineProtocol implementation. It uses ToolCallingLanguageModelProtocol to plan tools and records structured evidence in QueryEvidenceLedger.

question -> tool planning -> Wiki tools -> evidence ledger -> grounded synthesis

Default tools:

Tool	Purpose
read_wiki_index	Read page titles and summaries.
search_wiki	Fuse Wiki vector and full-text retrieval.
read_wiki_page	Read one generated Wiki page.
read_raw_object	Read non-citable supporting source context.

The default registry exposes this mode only when both Wiki indexes, an embedding model, and a tool-calling language model are available.

Results retain both their direct and original sources:

{
  "source": {
    "object_key": "wiki/pages/1-plant-cell.md",
    "object_type": "wiki",
    "origin": {
      "object_key": "raw/biology/plant-cell.pdf",
      "object_type": "pdf"
    }
  }
}

source.object_key can be passed to read_wiki_page, while source.origin.object_key identifies the document originally ingested into the knowledge base for audit or source lookup. Citations retain the same source structure.

The loop bounds model steps, tool calls, individual tool text, and total evidence. Final insights[].evidence_ids must reference citable ledger evidence; results and citations are derived only from selected evidence.

AgenticQueryEngine(
    max_steps=8,
    tool_budget=QueryToolBudget(
        max_tool_calls=8,
        max_tool_result_chars=8_000,
        max_evidence_chars=24_000,
    ),
)

Tool text can be truncated by these limits. For exhaustive questions over long pages, use search_wiki to retrieve multiple sections or explicitly increase ReadWikiPageTool(max_chars=...) and the total evidence budget. Unread page tails must not be treated as excluded evidence.