Cartograph
Somebody in the building has already solved this. The hard part was never the answer, it is knowing whose desk it is on.
- 6
- Source systems indexed with provenance
- 2
- Retrievers, lexical and semantic, always both
- 0
- Digits lost on the way into an embedding
The problem
Institutional knowledge is not missing, it is scattered: Slack, Jira, Drive, email, a database older than the team. And plain vector search loses the part that matters most, a clause number or an amount.
What we did
- 01
Index the systems that actually hold the knowledge, with provenance kept so an answer can name its source.
- 02
Search lexically and semantically at once. Exact match keeps the figures; the semantic half finds the thread that never used your words.
- 03
A semantic map of the estate, so a question about ownership resolves to a person.
- 04
Access stays exactly as the source systems set it. A search layer that quietly widens permissions is a breach.
Where it landed
A question about the organisation gets answered from the organisation, with the document and the owner attached. Knowing who to ask is usually faster than reading.
Built with
- Python
- PostgreSQL
- pgvector
- BM25 / hybrid rank
- Slack, Jira, Drive, Graph