Objectives
- Deliver precise answers with minimal tokens.
- Keep agents in flow without manual hunting.
- Blend public and private sources safely.
Search Loop
- Use session history to skip duplicate hits.
- Return only the top ~5k relevant tokens per read.
- Let agents adjust prompts while paging results.
Token Discipline
- Favor HTTP MCP mode for fast responses.
- Trim payloads before handing context to the agent.
- Track credits per endpoint to spot waste.
Source Coverage
- Index GitHub repos for code and docs.
- Add PDFs and Markdown libraries for gaps.
- Combine web search when private data misses.
Agent Experience
- Verify installs with a sample MCP prompt early.
- Surface usage cards and alerts in the dashboard.
- Keep role-based guardrails tight for teams.
Continuous Improvement
- Monitor credit logs after each feature launch.
- Re-run indexing when repos cross major releases.
- Capture wins and issues in a monthly changelog.