Threadmark

A team knowledge assistant that answered fluently before it could prove where answers came from.

FailedPythonVector SearchLLM

54%answers cited

0permission layers

2024year explored

The question

Could scattered project decisions become conversationally searchable?

I optimized retrieval breadth and response quality before defining citation and permission requirements.

The demo felt impressive but failed the trust test when users asked for authoritative evidence.

For workplace AI, a modest answer with evidence is more valuable than a brilliant answer without provenance.