agentic-search

name	agentic_search
description	Run a local-first research workflow by planning sub-queries, iterating retrieval, validating sources, and producing traceable evidence-backed conclusions.
metadata	{"level":"advanced"}

Agentic Search Skill

When To Use

Use this skill when the user asks for:

• multi-step research instead of a single lookup
• external verification for incomplete document evidence
• evidence-backed conclusions with explicit source traceability
• gap-filling loops (find missing facts, then re-search)

Do not use this skill for:

• direct extraction from a clearly scoped local document where one retrieval pass is enough
• purely stylistic rewriting tasks

Workflow

1. Define research objective and output contract

• Convert user request into explicit deliverables: key question, expected depth, and evidence contract.
• For project-document evidence, the contract is <evidence>chunk_id|p页码|o起止偏移</evidence>.

2. Build query plan

• Split into sub-queries by intent:
  • factual lookup
  • method comparison
  • recency update
  • risk/limitation check
• Prefer local document retrieval first.

3. Execute iterative retrieval

• Run local/document search first.
• If local evidence is insufficient, expand to scholarly search and then web search.
• Track each round with: query, source type, key hit, confidence.

4. Evaluate source quality

• Score sources by credibility, relevance, recency, and cross-source consistency.
• Downgrade claims that rely on a single weak source.

5. Synthesize answer with traceability

• Produce structured findings with explicit evidence anchors.
• When citing project-document evidence in prose, use <evidence>chunk_id|p页码|o起止偏移</evidence> instead of bracket citations or free-form page references.
• Separate confirmed facts, probable inferences, and unresolved gaps.

Runtime Contract

• Return conclusions that are evidence-backed and audit-friendly.
• For each key conclusion, attach at least one source anchor.
• For current-project document anchors, use the exact <evidence>chunk_id|p页码|o起止偏移</evidence> format.
• If evidence conflicts, explain conflict instead of forcing one conclusion.
• If evidence is still insufficient, output explicit follow-up search actions.

Progressive References

Read additional references only when needed:

• references/workflow_blueprint.md: full multi-round execution blueprint
• references/source_quality_rubric.md: scoring and conflict-resolution rubric
• references/output_schema.md: final structured output schema

Scripted Helpers

• scripts/evidence_aggregator.py: normalize and merge heterogeneous evidence records
• scripts/source_score.py: deterministic source scoring utility