Retrieval & Memory
Hybrid RAG pipelines and multi-tier session memory — automatic on every intent call.
Retrieval runs automatically inside the engine on every intent call — you never invoke it directly. Hybrid RAG pipelines with semantic chunking, metadata filtering, and cross-encoder re-ranking inject the right context into each call, backed by Knowledge Collections as the source of truth for what gets indexed.
Capabilities
Semantic Chunking
Documents are split at natural content boundaries — section headers, paragraphs — not fixed character offsets.
Hybrid Search
Dense vector search and BM25 keyword search run in parallel, combined via reciprocal rank fusion.
Cross-Encoder Re-Ranking
Top candidates are re-ranked against the query before reaching the model.
Bring Your Own Vector Store
Connect Pinecone, Weaviate, Qdrant, or pgvector via the adapter API (Enterprise).
Memory tiers
| Tier | Latency | Use case |
|---|---|---|
| Short-Term Buffer | < 1ms | Active single-session conversations |
| Episodic Memory | ~15ms | Multi-session agents — summaries injected at session start |
| Semantic Store | ~40ms | Knowledge-intensive tasks, RAG over large corpora |
Configure retrieval behavior per domain — see RAG configuration and Building a RAG Pipeline.