AI Software Development
Enterprise RAG implementation
Retrieval architectures designed to survive audit: source attribution, permission inheritance, and evaluation
- Duration
- 6 to 10 weeks
- Ladder stage
- Product
- NATIVE stages
- Integrate, Validate
The situation
When this engagement is the right one
Retrieval is easy to demonstrate and hard to govern. The three things that decide whether a knowledge system survives contact with an enterprise are permission inheritance, source attribution, and an evaluation set that includes the questions nobody wants asked.
You are probably seeing
- The assistant answers correctly for the pilot group and leaks across permission boundaries at scale.
- Answers have no citations, so nobody trusts them and usage stalls.
- Content is stale and there is no ingestion freshness signal.
What we do
The work, in the order it happens
- 01Map sources and their permission models before choosing a retrieval design.
- 02Inherit permissions at query time rather than filtering after retrieval.
- 03Attach source attribution to every generated answer.
- 04Build a retrieval evaluation set including adversarial and out-of-scope questions.
- 05Instrument freshness, coverage, and unanswered-query rate.
What you get, and keep
- A retrieval architecture document with the permission model stated.
- A production retrieval service with attribution.
- A retrieval evaluation set and harness.
- Freshness and coverage instrumentation.
Prerequisites
- A named content owner per source.
- An agreed answer for what happens when the system does not know.
Not included
- We do not ingest sources whose permission model is undefined.
- We do not build enterprise search as a replacement for content ownership.
Duration and price shape
6 to 10 weeks
From a published figure, scoped by source count and permission complexity.
Where this sits in the method
Commercially this is a Product engagement on the Proof, Product, Platform ladder.
FAQ
Questions we get asked
- RAG or fine-tuning?
- Different problems. Retrieval changes what the model knows now; fine-tuning changes how it behaves. Most enterprise cases start with retrieval.
- How do you handle permissions?
- Inherited from the system of record at query time. Post-filtering is a leak waiting for scale.
Related engagements
Work that usually sits either side of this
Next step
Ready to scope enterprise rag and knowledge systems?
Bring the pilots you already have running. The first call is a scoping conversation, not a pitch.