WARC + AI: Experimental Retrieval Augmented Generation Pipeline for Web Archive Collections.

More info:

• "WARC-GPT: An Open-Source Tool for Exploring Web Archives Using AI". Feb 12 2024 - lil.law.harvard.edu

• Features
• Installation
• Configuring the application
• Ingesting WARCs
• Starting the server
• Interacting with the Web UI
• Interacting with the API
• Visualizing Embeddings
• Disclaimer

• Retrieval Augmented Generation pipeline for WARC files
• Highly customizable, can interact with many different LLMs, providers and embedding models
• REST API
• Web UI
• Embeddings visualization

☝️ Summary

WARC-GPT requires the following machine-level dependencies to be installed.

• Python 3.11+
• Python Poetry

Use the following commands to clone the project and instal its dependencies:

git clone https://github.com/harvard-lil/warc-gpt.git
poetry env use 3.11
poetry install

☝️ Summary

This program uses environment variables to handle settings. Copy .env.example into a new .env file and edit it as needed.

cp .env.example .env

See details for individual settings in .env.example.

A few notes:

• WARC-GPT can interact with both the OpenAI API and Ollama for local inference.
  • Both can be used at the same time, but at least one is needed.
  • By default, the program will try to communicate with Ollama's API at http://localhost:11434.
  • It is also possible to use OpenAI's client to interact with compatible providers, such as HuggingFace's Message API or vLLM. To do so, set values for both OPENAI_BASE_URL and OPENAI_COMPATIBLE_MODEL environment variables.
• Prompts can be edited directly in the configuration file.

☝️ Summary

Place the WARC files you would to explore with WARC-GPT under ./warc and run the following command to:

• Extract text from all the text/html and application/pdf response records present in the WARC files.
• Generate text embeddings for this text. WARC-GPT will automatically split text based on the embedding model's context window.
• Store these embeddings in a vector store, so it can be used as WARC-GPT's knowledge base.

poetry run flask ingest

# May help with performance in certain cases: only ingest 1 chunk of text at a time.
poetry run flask ingest --batch-size 1

Note: Running ingest clears the ./chromadb folder.

☝️ Summary

The following command will start WARC-GPT's server on port 5000.

poetry run flask run
# Not: Use --port to use a different port

☝️ Summary

Once the server is started, the application's web UI should be available on http://localhost:5000.

Unless RAG search is disabled in settings, the system will try to find relevant excerpts in its knowledge base - populated ahead of time using WARC files and the ingest command - to answer the questions it is asked.

The interface also automatically handles a basic chat history, allowing for few-shots / chain-of-thoughts prompting.

☝️ Summary

Returns a list of available models as JSON.

Performs search against the vector store for a given message.

Uses an LLM to generate a text completion.

Returns RAW text stream as output.

☝️ Summary

WARC-GPT allows for generating basic interactive T-SNE 2D scatter plots of the vector stores it generates.

Use the visualize command to do so:

poetry run flask visualize

visualize takes a --questions option which allows to place questions on the plot:

poetry run flask visualize --questions="Who am I?;Who are you?"

☝️ Summary

The Library Innovation Lab is an organization based at the Harvard Law School Library. We are a cross-functional group of software developers, librarians, lawyers, and researchers doing work at the edges of technology and digital information.

Our work is rooted in library principles including longevity, authenticity, reliability, and privacy. Any work that we produce takes these principles as a primary lens. However due to the nature of exploration and a desire to prototype our work with real users, we do not guarantee service or performance at the level of a production-grade platform for all of our releases. This includes WARC-GPT, which is an experimental boilerplate released under MIT License.

Successful experimentation hinges on user feedback, so we encourage anyone interested in trying out our work to do so. It is all open-source and available on Github.

Please keep in mind:

• We are an innovation lab leveraging our resources and flexibility to conduct explorations for a broader field. Projects may be eventually passed off to another group, take a totally unexpected turn, or be sunset completely.
• While we always have priorities set around security and privacy each of those topics is complex in its own right and often requires grand scale work. Experiments can sometimes initially prioritize closed-loop feedback over broader questions of security. We will always disclose when this is the case.
• There are some experiments that are destined to become mainstays in our established platforms and tools. We will also disclose when that’s the case.

☝️ Summary