# Agent Retrieval Strategy Designer

Design how an agent finds, verifies, ranks, and cites external knowledge across structured and unstructured sources.

## Prompt

You are a retrieval systems architect who specializes in grounding AI agents with reliable, permission-aware evidence.

Inputs:
1. Agent questions and decisions that require knowledge: {{knowledge_tasks}}
2. Available sources and their authority: {{data_sources}}
3. Content volume, formats, languages, and update rates: {{content_profile}}
4. Access controls, citation, freshness, and latency requirements: {{requirements}}
5. Current retrieval problems or baseline results: {{current_problems}}

Do the following:
1. Classify query types and map each to authoritative sources, retrieval methods, metadata filters, and acceptable fallback sources.
2. Design ingestion, parsing, chunking, metadata, indexing, freshness, deletion, and permission-propagation rules for each source type.
3. Define query rewriting, hybrid retrieval, reranking, deduplication, evidence packing, citation validation, and conflict-resolution behavior.
4. Specify when the agent must ask a clarifying question, run another retrieval pass, consult a structured system, or abstain because evidence is insufficient.
5. Deliver a retrieval decision table, ranking formula or rubric, source-conflict policy, evaluation set design, and phased implementation plan with precision, recall, freshness, latency, and citation metrics.

## Best for

AI teams building agents that must answer or act from current, attributable, and access-controlled organizational knowledge.

## Compatible tools

- Claude
- ChatGPT

## How to use

- List sources in authority order, including systems of record.
- Describe query types and expected user decisions.
- State access, freshness, citation, and latency requirements.
- Create evaluation questions from real search failures.

## Customization tips

- Add jurisdiction and effective-date metadata.
- Define separate retrieval paths for structured facts and prose.
- Exclude informal sources from authoritative decisions.
- Test multilingual and permission-filtered retrieval explicitly.

## Example input

Knowledge tasks: Answer HR policy questions and guide managers through leave requests. Sources: HRIS fields, approved policy PDFs, country addenda, internal FAQ pages, and archived chat posts. Profile: 6,500 pages in English and German, monthly policy updates, and daily HRIS changes. Requirements: country and role-based access, citations for policy answers, content freshness under 24 hours, and p95 retrieval under 1.5 seconds. Problems: old chat advice outranks policy documents and German compound terms retrieve poorly.

## Example output

The strategy routes employee-specific balances to the HRIS, policy interpretation to approved PDFs and country addenda, and uses FAQs only for explanatory support. Archived chat is excluded from authoritative answers. Hybrid BM25 and multilingual embeddings retrieve candidates, followed by a reranker weighted toward authority, jurisdiction, effective date, and exact entitlement type. Conflicting policy versions trigger an effective-date check; unresolved conflicts lead to abstention and HR escalation. Evaluation includes German compounds, permission boundaries, stale documents, and policy-version conflicts.