Solidchain

Caution: currently this is just a small testbed, and should not be used or taken for a real app/reference.

This is a web-based browser interface for searching with Large Language Models (LLMs). Built with SolidJS and LangChain, it allows you to index a collection of documents (such as markdown files, PDFs, webpages, chat logs, etc.) and search using a dialogue based Q/A with natural language. Methods such as chain of thought (CoT) reasoning, and agent based reasoning (e.g. MRKL & ReAct) can be used depending on the use case.

Installation

This application has two components, a frontend built with SolidStart, and an api layer built with FastAPI and LangChain.

Web

To run the webapp:

cd apps/web
pnpm install
pnpm dev

Web Extension

To run the extension locally:

cd apps/web-extension
pnpm install
pnpm dev

To view this extension (with HMR), go to chrome://extensions and click Load unpacked choosing the build/chrome-mv3-dev directory. See the Plasmo Docs for more detail.

API

To run the api, first create a virtual environment. We provide a sample environment.yml file to use with conda environments:

cd apps/api
conda create -f environment.yml
conda activate -n solidchain

Now run the api (in dev) with:

uvicorn src.solidchain.main:app --host ::1 --port 8000 --reload

The documentation (OpenAPI) can be found at localhost:8000/docs.

To run the initial database migration (initializing your database), run:

alembic upgrade head

In development, when creating changes to the schema, create a new migration with:

alembic revision --autogenerate -m "Add column X and index Y to model Z"

Deployment

We provide a sample docker-compose.yml file in the root of the project. You can use this by running:

docker-compose up

This runs a local (postgres) database with a volume postges-data, as well as the pgadmin4 UI. In order to setup a connection, go to localhost:5050/pgadmin4 and register a server. The local defaults (for convenience) are:

DATABASE=postgres
USERNAME=postgres
PASSWORD=postgres
PORT=5432

The host (locally) can be found by running:

docker inspect CONTAINER_ID  | grep IPAddress.

TODO

Possible extensions

• allow for local inference: e.g. flexgen
  • llama/opt
• web extension:
  • capture users clipboard/current page as context
  • extension bookmarking (upload html page to collection)
• discord extension
• visualization
  • stable diffusion (image generation toolkit + auto prompt engineering)
  • llm call visualization (tracing)
• starter examples (email, mermaid graphs, terminal autocompletion)