Retrieval bench
Where the agent’s evidence comes from when it hasn’t spent a cent: 11 keyless indexes plus keyed providers when configured, document parsing for the files the web is actually made of, and honest reporting when a door is closed.
Reference: dr-nib/backend/src/retrieval/.
Free indexes (no key, no signup)
GDELT (paced), Wikipedia, OpenAlex, Semantic Scholar, Crossref, SEC EDGAR, Stack Exchange, Hacker News, Polymarket, arXiv, plus a self-hosted SearXNG instance. Tavily/Exa slots accept keys when configured. Queries are keyword-ified per provider; provider failures are tagged by name so the trail shows which source failed, not just that something did.
Documents, not just pages
Fetched files are routed by magic bytes (not extensions) through real parsers, 10MB cap, single download:
| Format | Parser |
|---|---|
| text extraction | |
| DOCX | document conversion |
| XLSX | spreadsheet parsing |
| PPTX | slide XML extraction |
Bot-block honesty
When a site blocks machine reading, the run records blocked-bot-check with the provider named — reported, never worked around with evasion tricks. A thin-retrieval warning is itself evidence: it tells the data stage to reach for primary APIs, sandbox compute, or paid sources instead.
How it feeds the loop
Search fans each sub-question across the bench; fetch opens ranked pages; JEV scores every source for relevance and trust before it joins the scorable set. Paid content that clears scoring joins on equal footing — paying never buys credibility. When the bench comes back thin, the loop’s data stage escalates: model-directed API calls, chain RPC reads, sandbox crunching, then spend tools in that order.