Cognitive Extraction Pipeline

Neural Entity Extraction

Tellodb integrates a local BERT-based model for Named Entity Recognition (NER). During ingestion, episodic text is scanned to identify core entities without requiring external LLM calls.

Entities are classified into standard categories (Person, Organization, Location, Miscellaneous), allowing for precise scoping and relationship mapping.

• PER: Identities and individual actors.
• ORG: Companies, teams, and institutions.
• LOC: Geographic context and movement history.
• MISC: General artifacts and specific concepts.

Autonomous Relationship Discovery

Extracted entities are passed through a heuristic relationship engine that automatically constructs knowledge graph triples.

For example, detecting a Person and an Organization in the same context can trigger an `associated_with` edge, while two people can trigger a `knows` edge.

Implicit Preference Detection

Tellodb identifies sentiment and preference signals (love, hate, prefer, favorite) automatically. When detected, the memory is elevated to a `Preference` kind, exempting it from standard time-decay policies to ensure core user identity persists.