
RAG & Retrieval
Answers your users can act on
Most document systems answer plausibly and wrongly. So the work starts in your index.
- What gets indexed, and in what pieces
- What the model may say when the index has nothing
- A source your users can open on each answer
- A suite that scores answers against your documents
Example record
An answer you can check
Each answer carries the passages it came from. Your users can check it themselves.
When the index has nothing, the answer says so.
- Done
Question scoped
claims policy, version in force
- Cited
Passage from the current policy
opens at the clause
- Excluded
Superseded policy
filtered out by version
- To a person
Overseas claims
the index has nothing on it
Why a plausible answer is the dangerous one
A wrong answer that sounds right costs more than no answer, because nobody checks it.
Chunks cut through the middle of a table
A superseded policy ranked above the current one
An answer stitched from 2 documents that disagree
A confident reply when the index had nothing
Capabilities
Most of it happens before the question
The model is the last step. Your answers depend on the 4 before it.
Model
What we fix, and what you decide
What we fix
- F1An evaluation suite built from your documents
- F2A citation on each answer
- F3A defined answer for nothing found
What you decide
- C1Which documents go in, and who may see them
- C2Vector store and hosting
- C3When an answer goes to a person
Next step
Send us 3 questions it gets wrong.
We trace each one to the index. You decide which fix to fund first.

