recon
- python
- crewai
- sqlite
- embeddings
Ask a «deep research» tool for a report on any topic and it’ll hand you something that sounds flawless. The problem is that sounding flawless and being true aren’t the same thing, and almost none of these tools tells you which facts to trust.
Recon splits that work in two. First, several agents research the topic in parallel, each from a different angle. Then —and this is what changes things— a separate agent acts as verifier: it extracts each claim, looks for its source, and tags it. Verified, partially verified, unsourced, or contradicted. Only then is the report written, with the tags in plain sight.
Splitting researching from verifying isn’t an implementation detail, it’s the idea. The one researching is busy searching; the one verifying has no job but to doubt. And you end up reading a report that tells you what to trust and what not to, instead of one that hides it.
How it’s built — Python · CrewAI · SQLite + FTS5 · local ONNX embeddings