Agents & Grounded Content
Research Agent
A plan → search → evaluate → synthesise pipeline that cites evidence or admits it has too little.
The Problem
Agent demos usually stop at “call a tool and summarise”. Production agents need cost accounting, evidence grading and a way to refuse.
How It Works
- Decompose the question into sub-questions.
- Call a search tool (cached, with retries) and score each result: 0.5 relevance + 0.3 source credibility + 0.2 recency.
- Accept evidence at ≥ 0.35, keep the top three per section, then re-validate every citation before it reaches the report.
Technically Interesting
The pipeline reports its own tool calls, cache hits, estimated tokens and confidence, and revalidates citations independently of the synthesiser.
Why It Matters
This is the engineering around the model — grounding, budgets and honest failure — that decides whether an agent is trustworthy in an enterprise.
Scope & honesty. The recorded run uses a bundled synthetic corpus, mock keyword retrieval and a template synthesiser (no live LLM call). The LLM client is a pluggable interface.