Decomposed Local Research

Why decompose

A long agentic loop can exceed a local model's tool-call or context budget even when it can summarize and compare evidence well. Separate deterministic retrieval from bounded language work so each model call has one focused purpose.

Architecture

1. A deterministic retriever reads the supplied source and retrieves the selected evidence URLs into an evidence pack.
2. A bounded model pass summarizes the source and extracts checkable claims.
3. A bounded verification pass labels each claim supported, contradicted, unsupported, or overstated against the evidence pack.
4. A bounded synthesis pass writes the verdict using only the prior stages; it introduces no new facts.

Procedure

• The orchestrator chooses the source and only critical evidence URLs.
• Keep evidence packs small enough for the target context window; split or summarize source material when needed.
• If needed evidence was not retrieved, add it and rerun retrieval rather than asking the model to guess.
• Use discovery-capable research or a human review for open-ended source discovery and reliability-critical work.

Pitfalls

• Handing a local model a monolithic fetch-and-write agent loop.
• Calling a claim verified when the evidence pack does not cover it.
• Treating a supplied-evidence workflow as web discovery.

Verification checklist

• Evidence was selected and retrieved before claim verification.
• Unsupported claims are labeled rather than filled in.
• The final synthesis is traceable to the evidence pack.