/pacha-ai

Connect your private data to LLMs

Primary LanguagePLpgSQL

Patcha DDN

Connect realtime data to your AI

What is it?

Pacha is an AI tool that does retrieval of context for a natural language query using a SQL interface. Pacha is specially built to work with Hasura DDN for authorized multi-source querying.

You would use Pacha with your favourite LLMs to generate grounded responses in your AI apps/agents/chatbots.

Getting Started

Prerequisites

  • Atleast Python version 3.12
  • Access to OpenAI or Anthropic.
  • A postgres database you want to try Pacha out on.

Install Python Dependencies

  • Install Poetry
  • Run poetry install to install Python dependencies.

Setup Hasura DDN

Note: You can skip this step if running Pacha directly against Postgres instead of Hasura DDN.

  • Create a Hasura account at hasura.io/ddn
  • Scaffold a local Hasura setup on a postgres database like this:
poetry run ddn_setup -c <postgres connection string> --dir ddn_project
  • The above generated metadata is where you would configure row / column access control rules for your data.
  • Start a local Hasura engine with:
docker compose -f ddn_project/docker-compose.hasura.yaml up -d

Running Pacha

examples/chat_with_tool.py is a CLI chat interface that uses Pacha with Anthropic.

ANTHROPIC_API_KEY=<api-key> poetry run chat_with_anthropic -d ddn -u <DDN SQL URL> -H <header to pass to DDN>

Example:

ANTHROPIC_API_KEY=<api-key> poetry run chat_with_anthropic -d ddn -u http://localhost:3000/v1/sql -H 'x-hasura-role: admin'

You can also run Pacha with OpenAI:

OPENAI_API_KEY=<api-key> poetry run chat_with_openai -d ddn -u <DDN SQL URL> -H <header to pass to DDN>

Customizing

Running against a custom SQL backend

If you want to run against a custom SQL backend that's not Hasura DDN or Postgres, you can implement the DataEngine class in pacha/data_engine, and pass that to the Pacha SDK. See usage here.

You can see example Postgres implementation of DataEngine is in pacha/data_engine/postgres.py.

Running against a custom LLM

You can run Pacha against any LLM that supports function/tool calling. You can see the examples in examples/chat_with_tool_anthropic.py and examples/chat_with_tool_openai.py.