01Overview
Synapse stitches together a hybrid retrieval pipeline (BM25 + dense vectors + reranker) over a polyglot corpus of code, docs, RFCs, and chat. The IDE plugin streams results inline as engineers type code comments or commit messages.
The ingestion pipeline is a Kafka-fed Python worker that performs structure-aware chunking — code by AST, prose by semantic break points — before embedding through a fine-tuned bge-large model.
02The Problem
A 400-engineer org had answers everywhere and findability nowhere. Tribal knowledge was a tax on every onboard.
03Approach
- Built structure-aware chunkers per content type so retrieval respects natural unit boundaries.
- Designed a hybrid retriever with reciprocal-rank fusion, then a cross-encoder reranker on the top 50 candidates.
- Shipped an LSP-style plugin that streams suggestions through Server-Sent Events with sub-200ms TTFT.
04Outcome
Median time-to-first-answer for new hires dropped from 18 minutes to 40 seconds.
70% of weekly active users keep the IDE plugin enabled past the 30-day mark.
Now powering internal search for three subsidiary products.