InfoBridge AI
A 55-million-object RAG assistant that stays current
An enterprise wanted employees to ask natural-language questions across policy manuals, contracts and internal wikis. The corpus was large, changed daily, and answers had to be citable.
The challenge
- Tens of millions of chunks across 39 collections with different access rules.
- Documents changed constantly; full re-vectorisation was too slow and too expensive.
- Median latency had to stay low under concurrent load from thousands of employees.
What we built
Self-hosted vector platform
A Weaviate cluster inside the client's own environment, with RQ-8 quantisation giving roughly 4x compression while retaining 85–92% recall on the evaluation set.
Delta updates, not rebuilds
A Kafka-based change-data-capture pipeline listens to the document stores and applies only the deltas, so the index is minutes behind the source rather than a day.
Cache the obvious
Memcached fronts frequently asked queries; repeated questions never touch the vector store or the model.
Measure retrieval first
A golden question set and recall metrics drove chunking and hybrid-search decisions before any prompt work began.
What we learned
- Most answer failures were retrieval failures; fixing chunking beat prompt engineering every time.
- Quantisation is a cost lever worth measuring per collection rather than globally.
- The CDC pipeline paid for itself in the first month of avoided re-indexing.
Want something like this?
Most of these projects started as a two-page brief. Send us yours.