Typed Agentic RAG with Pydantic AI

This Streamlit app answers questions from uploaded PDFs or a documentation URL.

Every response is a validated Answer object with exact source quotes, chunk IDs,

a confidence score, and an answered decision. If retrieval is too weak, the app

refuses before calling the language model.

Features

• Pydantic AI Agent, RunContext, and dependency injection
• A typed retrieve tool with source metadata and cosine scores
• Pydantic models for answers, citations, and retrieval evidence
• Exact quote checks against indexed chunks after model output validation
• A deterministic refusal gate for out-of-corpus questions
• OpenAI or Anthropic answer models
• OpenAI embeddings with a local hashing fallback for Anthropic-only setups
• A session-scoped NumPy vector store with no database service

How it works

1. rag.py extracts PDF or web text, splits it into overlapping chunks, embeds

the chunks, and stores normalized vectors in memory.

2. agent.py injects the vector store through RagDependencies. The Pydantic AI

agent must call the typed retrieve tool before producing an Answer.

3. A preflight search compares the best cosine score with the refusal threshold.

Low scores return answered=False without an LLM request.

4. For an answered response, each citation must match a stored source, chunk ID,

and verbatim quoted span. An invalid or missing citation becomes a refusal.

5. app.py renders the answer, confidence, citations, or refusal state.

When OPENAI_API_KEY is available, Auto mode uses Pydantic AI's OpenAI

Embedder with text-embedding-3-small. With only ANTHROPIC_API_KEY, Auto mode

uses the local hashing backend because Anthropic has no embeddings API. The local

backend is best for keyword-oriented demos. Select OpenAI embeddings for semantic

retrieval across paraphrases.

Prerequisites

• Python 3.12 or newer
• An OpenAI API key or an Anthropic API key

Setup

From the repository root:

cd rag_tutorials/agentic_typed_rag_pydanticai
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env

Add one key to .env:

OPENAI_API_KEY=your-key

or:

ANTHROPIC_API_KEY=your-key

The default answer models are openai:gpt-5.2 and

anthropic:claude-sonnet-4-6. Change the model field in the sidebar or set

RAG_MODEL to another Pydantic AI model string.

Run

From rag_tutorials/agentic_typed_rag_pydanticai:

streamlit run app.py

Upload one or more PDFs, optionally add a docs URL, and select **Build knowledge

base**. Ask an in-corpus question to see a cited answer. Then ask about an

unrelated topic to see the refusal state.

Tests

The deterministic suite uses Pydantic AI's TestModel, so it makes no provider

requests:

python3 test_typed_rag.py

Files

agentic_typed_rag_pydanticai/
├── app.py
├── agent.py
├── rag.py
├── test_typed_rag.py
├── requirements.txt
├── .env.example
└── assets/screenshot-placeholder.svg

Licensed under Apache-2.0 as part of awesome-llm-apps.