langchain-ai-app

A demo that answers natural-language questions about onchain Uniswap swap data by letting a LangChain agent query an SQD (Subsquid) GraphQL squid. It has two parts: a Python Flask backend that runs the agent, and a Next.js frontend that sends prompts to it.

This is an example project from SQD (Subsquid Labs). It shows one way to put a language-model agent in front of data indexed with the SQD stack.

What it does

A user types a prompt (for example, a question about Uniswap swaps). The frontend sends the prompt to the backend's HTTP endpoint. The backend runs a LangChain agent backed by OpenAI that has three tools:

• introspect the GraphQL schema of a squid and summarize it as text,
• run a GraphQL query against that squid,
• sort the returned results by timestamp or block number.

The agent uses these tools to build and execute a query, then returns a text answer. The squid it queries is a Uniswap swaps squid served at a hosted GraphQL endpoint.

Repository layout

• llama-backend/: Flask app (main.py) exposing the agent over HTTP. The agent and its tools are defined in base/base.py. Uses langchain, llama-index, and the OpenAI API.
• llama-frontend/: Next.js (App Router) app. The prompt form (src/app/_components/emoji-form/index.tsx) POSTs to the backend's /post_json endpoint. The frontend was adapted from the pondorasti/emojis template, so some files still carry that template's naming.

Backend

Endpoints

• GET / and GET /about: return a placeholder JSON message.
• POST /post_json: accepts JSON { "prompt": "..." } with Content-Type: application/json and returns { "text": "..." } with the agent's answer.

Requirements

• Python 3.11
• An OpenAI API key

The backend reads the OpenAI key from a variable in the source (main.py). Set your own key there or wire it to an environment variable before running. Do not commit a real key.

Install and run

cd llama-backend
pip install -r requirements.txt
python3 main.py

This starts the development server on port 5000 (override with the PORT environment variable). For production, the included Procfile runs the app with Gunicorn:

web: gunicorn main:app --timeout 900

Key dependencies (see requirements.txt): Flask, flask-cors, langchain, llama-index, llama-hub, openai, requests, python-dotenv, gunicorn.

Frontend

A Next.js App Router application that renders a prompt box and shows the agent's reply with a typewriter effect. The request target is set in src/app/_components/emoji-form/index.tsx; change it to point at your own backend deployment.

Install and run

The project includes both a bun.lockb and a package-lock.json, so you can use Bun or npm.

cd llama-frontend
npm install        # or: bun install
npm run dev        # starts Next.js on port 3002

Other scripts (package.json): build, start, lint, prisma:studio, prisma:push.

Note: the frontend retains parts of the upstream emoji-generator template (Prisma schema, Replicate, Vercel Blob/KV references). Those are not required for the prompt-to-backend flow described above and can be removed or ignored for that use case.

How this fits into SQD

SQD is an open data platform for Web3. Squids are indexers built with the Squid SDK that expose indexed onchain data, including over a GraphQL API. This project queries one such squid (Uniswap swaps) and uses a language-model agent to translate natural-language prompts into GraphQL queries against it.

• SQD docs: https://docs.sqd.dev
• Squid SDK docs: https://docs.sqd.dev/en/sdk
• SQD: https://sqd.dev

Related projects used here

• LangChain: agent and tool framework
• LlamaIndex: used for an alternative agent setup in base/base.py
• Flask: backend web framework
• Next.js (App Router): frontend framework