Case study 04
Retrieval, brought in-house
A $5k/month third-party AI dependency, replaced with a custom vector store and crawler the product fully owns.
Outcome
$60k+/yr saved
- Removed
- $5k / month
- Saved
- $60k+ / year
- Ownership
- End to end
Context
Our AI voice-agent SaaS leaned on a third-party vendor's managed knowledge base for the retrieval behind every call - $5k a month, and climbing with every knowledge base our customers created. It sat at the center of the product. My role: make the build-vs-buy case honestly, then ship the replacement.
The problem
The cost was the visible problem; ownership was the real one. Retrieval quality, indexing cadence, and roadmap were all rented, and the bill scaled with the customer count rather than the value. Every improvement the product needed was a feature request to someone else's backlog, billed monthly.
The outcome
For the business: $60k+ per year in recurring vendor cost eliminated, and the roadmap for a core capability moved in-house. For the system: a retrieval stack - vector store, crawler, indexing pipeline - that the product fully owns and can change at will.
The approach
Build over buy - argued in writing, not by instinct. The workload was well understood and stable, which is exactly the case where owning the implementation beats renting it. I put the trade-offs on paper before writing code.
Our own RAG pipeline on an OpenAI vector store, in place of the vendor's managed knowledge base. Owning ingestion, chunking, and indexing meant retrieval quality and cadence became ours to tune - not a line item that grew with every knowledge base a customer added.
A crawler that builds a knowledge base from a single URL. A customer pastes a website link; the system crawls it, compiles a full knowledge-base document, and hands it to the vector store to ingest - no manual document assembly in the loop.
Built with