Our own tooling
One architectural question, priced 2 ways
An index of how a codebase relates cut the cost of an architectural question by a factor of 25.
Before
What was true before
A single architectural question on a mid-sized codebase meant reading to find out. Grep a name, open what matched, read the next name found there, open that too. Everything opened stayed in the context window and was re-sent to the model on every turn, so the question cost around 10,000 tokens before any work began. A sweep across an unfamiliar repository ran to between 25,000 and 50,000.
After
What we did
We indexed the relationships once instead of rediscovering them each time. Documents, concepts and code become nodes. References and dependencies become edges. One workspace came to roughly 700 nodes and 1,000 edges. Extraction runs on a local open-weights model at no API cost, or on a hosted model for a fraction of a cent per document, and refreshes as the code changes. The agent stops reading to find out and starts asking: a relationship query returns the part of the graph that answers the question.
Measured
Results
As measured on our own tooling. Not a client result.
- tokens per architectural question, against ~10,000 before
- ~400
- to return the answer
- Under 0.3 seconds
- less context for the same question
- 25x
- nodes and 1,000 edges indexing one workspace
- 700
Engagements
The services behind it
Engagement
Where this work sits
Next step
Yours would be the next one.
Tell us where it breaks, and you get a written answer on what breaks first and what it costs.