Skip to main content

​
What is TypeAgent?

TypeAgent is a Python library that implements Structured RAG (Retrieval Augmented Generation) for conversation-based knowledge processing. It enables you to ingest messages, emails, transcripts, and other conversational content, then query that content using natural language with high precision and recall.
TypeAgent is a Pythonic translation of TypeAgent KnowPro from TypeScript, developed by Microsoft Research.

​
Why TypeAgent?

Traditional RAG systems rely primarily on semantic similarity search, which can miss important context or return irrelevant results. TypeAgent takes a different approach:
  • Structured Knowledge Extraction: Automatically extracts entities, topics, actions, and relationships from your content
  • Multi-Index Architecture: Uses six specialized indexes for different query patterns (semantic refs, properties, timestamps, embeddings, related terms, and conversation threads)
  • Parallel Search: Executes queries across multiple indexes simultaneously and fuses results intelligently
  • Graceful Degradation: Works with basic extraction when AI models are unavailable
  • Provider Agnostic: Supports OpenAI, Azure OpenAI, and other providers via pydantic-ai

​
Key Benefits

High Precision

Structured knowledge extraction ensures accurate retrieval of relevant information from your conversations

Flexible Storage

Choose between in-memory storage for fast prototyping or SQLite for persistent data

Easy Integration

Simple async API with just two main operations: add_messages_with_indexing() and query()

Production Ready

Built on proven TypeChat and pydantic-ai foundations with comprehensive type annotations

​
How It Works

1

Ingest Content

Add messages, emails, or transcripts to a conversation using add_messages_with_indexing()
2

Extract Knowledge

TypeAgent automatically extracts entities, topics, actions, and relationships using AI models
3

Build Indexes

Content is indexed across six specialized indexes for fast, accurate retrieval
4

Query Naturally

Ask questions in natural language using the query() method
5

Get Answers

Receive AI-generated answers grounded in your indexed content with proper citations

​
Core Architecture

TypeAgent’s architecture consists of four main components:

​
1. Storage Layer

Dual-implementation pattern with identical APIs:
  • MemoryStorageProvider: Fast in-memory storage for prototyping
  • SqliteStorageProvider: Persistent SQLite-backed storage for production

​
2. Knowledge Extractor

Multi-mode content processing:
  • Basic Mode: Rule-based extraction from metadata and structure
  • AI Mode: LLM-powered entity and relationship extraction
  • Hybrid Enhancement: Combines both approaches for best results

​
3. Query Pipeline

Structured RAG system with:
  • Natural language query translation
  • Parallel multi-index search
  • Result fusion and ranking
  • Context-aware answer generation

​
4. Integration Layer

Unified conversation interface:
  • ConversationBase: Main entry point for all operations
  • create_conversation(): Factory function for easy setup
  • Async-first API design

​
Use Cases

Ingest your email archive and ask questions like “What decisions were made about the product launch?” or “Who discussed the budget in Q4?”
Process meeting transcripts and query action items, decisions, or topics discussed by specific participants
Index support conversations and quickly find similar issues, resolutions, or expertise by team members
Analyze podcast transcripts, interviews, or research conversations to extract insights and themes

​
What’s Next?

Installation

Get TypeAgent installed with pip, uv, or poetry

Quickstart

Build your first TypeAgent application in minutes
TypeAgent sends input to LLM providers for processing. Do not use it to index confidential information unless you control the LLM endpoint.

​
Project Status

TypeAgent is an experimental prototype under active development. The API is stabilizing but may change in future releases. It is suitable for:
  • Research and experimentation
  • Proof-of-concept applications
  • Non-confidential data processing
  • Learning Structured RAG patterns
For production use with sensitive data, ensure you use a private LLM deployment.

Build docs developers (and LLMs) love

Get started for freeTalk to us