Cognee

The core cognee.cognify() function processes raw text through chunking, entity extraction, relationship identification, and graph storage in a single call. Each stage is a composable Python task that can be swapped or extended, letting you customize behavior without rewriting the pipeline. This makes the simplest case (ingest a PDF and query it) trivially easy while keeping advanced customization within reach.

Cognee stores every ingested entity in both a graph database (Neo4j, Kuzu, or NetworkX) and a vector store (Qdrant, LanceDB, pgvector, Weaviate, or Milvus). Retrieval can combine graph traversal for relational queries with vector similarity for semantic search, giving you flexibility to answer different question types from the same knowledge base. This dual representation is the key technical differentiator versus pure-vector RAG frameworks.

You can define domain-specific schemas as Pydantic DataPoint subclasses — for example, a Patient class with fields for diagnoses, medications, and providers. The pipeline uses these schemas to guide structured extraction from documents, producing typed entities rather than generic strings. This is critical for regulated domains where extracted data feeds downstream systems requiring strict typing and validation.

Cognee integrates with 30+ LLM providers through LiteLLM, including OpenAI, Anthropic, Google, Azure, AWS Bedrock, Groq, Ollama, and self-hosted models. You can mix providers across the pipeline — for example, using a cheaper model for chunking and a stronger model for entity extraction. This flexibility avoids vendor lock-in and lets teams optimize cost vs quality per pipeline stage.

The hosted cloud tier provides a dashboard for graph exploration, pipeline monitoring, and data source management without requiring teams to operate Neo4j or vector infrastructure themselves. It includes visualization of the knowledge graph, ingestion job tracking, and team collaboration features. This bridges the gap for teams that want Cognee's capabilities without the DevOps burden of self-hosting the full stack.