Overview
CORE’s search is fundamentally different from traditional RAG. Instead of treating every query the same way, CORE first understands what kind of question you’re asking, then applies the optimal search strategy for that query type. This makes searches 3-4x faster and significantly more precise.The Search Pipeline
When you ask the CORE Agent a question or search your memory:Query Classification
CORE analyzes your query to determine the query type (aspect, entity, temporal, exploratory, or relationship), relevant time range, and which aspects to prioritize.
Strategy Selection
Based on the query type, CORE selects the search strategy:
- Aspect Query → Filter by aspect, then vector search
- Entity Lookup → Graph traversal from entity node
- Temporal Query → Time-based filtering first
- Exploratory → Recent session summaries
- Relationship Query → Multi-hop graph traversal
Search Execution
CORE runs the search using a hybrid of three methods:
- Vector search for semantic similarity (“what’s my approach to error handling?” finds related content even without exact keywords)
- BM25 keyword search for precise term matching (“Stripe integration” ranks documents with those terms highly)
- Graph traversal for relationship discovery (follows entity connections across multiple hops)
Relevance Ranking
Results are scored based on semantic relevance, recency, entity importance, label/topic match, and aspect priority. More recent facts rank higher while historical facts remain accessible.