doc·chat

What happens between the file and the answer

Short version for the person paying, then the engineering underneath.

In one sentence

The business effect

Your staff stop reading a 200-page manual to answer one question, and every answer they get carries the page it came from, so it can be checked in seconds instead of trusted blindly.

The analogy

A language model on its own is a colleague with a very good memory who has never read your documents. This system is the same colleague sitting with your file open, only allowed to answer from the page in front of them, and required to point at it.

Measured difference on a fixed question set: 15% correct without the document, 100% with it.

The pipeline

Timings are measured on this server right now, not written by hand.

1
Read the file

2
Split into passages

3
Compute embeddings locally

4
Search twice, merge

< 30 ms
5
Answer, or refuse

Two decisions worth explaining

Why it says “evidence”, not “confidence”

Price lists

Limits, honestly

Privacy