Four retrieval signals, one answer, half a second
A real-time agent-assist engine that turns noisy live-call transcripts into knowledge-base answers while the customer is still talking. A small language model first decides whether the turn deserves a search at all — chatter and filler get an explicit no-query verdict rather than a bad answer — then compiles the messy speech into a structured query with product and concept hints validated against a closed vocabulary.
Retrieval fuses four disagreeing signals: lexical match and dense embeddings combined by reciprocal-rank fusion, then a cross-encoder pass, then an instruction-tuned reranker that scores each document against the actual support task rather than generic similarity. Absolute and relative score floors sit under all of it, session memory carries confirmed topic keywords across turns and decays them the moment the conversation shifts, and account numbers are redacted before any model sees a word. When the floors are not met, the agent simply is not interrupted.
Behind it, a thousand-document knowledge base — web content, PDFs, spreadsheets, image-only scans read by a vision model — is distilled into scored, agent-facing cards. The latency budget taught the real lesson: the hard part was never any single retriever. It was making four signals that disagree resolve into one defensible suggestion, in half a second, on hardware the customer controls.
More detail
The four-signal fusion path, with confidence gating before generation.