ACMI collection chat

Uses LangChain, LangServe, and GPT-4 to chat with the ACMI Public API collection.

Run it on your computer

• Build the virtual environment: make build-local
• Setup your OpenAI API Key: cp config.tmpl.env config.env
• Install direnv if you'd like to load API Keys from the config.env file: brew install direnv
• Load the environment direnv allow
• Start chatting on the command line: make up-local
• Start chatting in a web browser: make server-local and visit: http://localhost:8000/playground
• See the API server documentation: http://localhost:8000/docs

Or if you have your own Python environment setup:

• Install the dependencies: pip install -r requirements.txt
• Start chatting on the command line: python chat.py
• Start chatting in a web browser: python api/server.py and visit: http://localhost:8000/playground
• See the API server documentation: http://localhost:8000/docs

Or to run the API server using Docker:

• Build: make build
• Run: make up
• Visit: http://localhost:8000
• Clean-up: make down

Environment variables

Optional environment variables you can set:

• DATABASE_PATH - set where your Chromadb vector database is located
• COLLECTION_NAME - the name of the Chromadb collection to save your data to
• PERSIST_DIRECTORY - the name of the directory to save your persistant Chromadb data
• MODEL - the large language model to use (defaults to OpenAI gpt-4o)
• EMBEDDINGS_MODEL - the embeddings model to use (defauls to OpenAI text-embedding-ada-002)
• REBUILD - set to true to rebuild your Chromadb vector database
• ALL - set to true to rebuild with the entire ACMI Public API collection

When using Docker:

• CACHE_URL - the URL to your pre-generated embeddings database e.g. https://example.com/path/

The scripts/entrypoint.sh script will look for a file named works_db_chat.tar.gz at the CACHE_URL to process.

To create that file, generate your embeddings database locally and then run: tar -cvzf works_db_chat.tar.gz works_db

Open source LLM using Ollama

To use an open source model on your computer with Ollama:

• Install ollama with: brew install ollama
• Start the ollama server with: OLLAMA_HOST=0.0.0.0:11434 ollama serve
• Pull an open source model: ollama pull llama3
• Set MODEL=llama3
• Set LLM_BASE_URL=http://<YOUR_IP_ADDRESS>:11434
• Start the chat: make up

Re-build all collection items

By default we only build the first ten pages of the Public API into the vector database, but if you'd like the build the entire collection:

• Add ALL=true to your config.env
• Then either delete your persistant directory or also add REBUILD=true to your config.env
• Rebuild the app: make build-local

Run it on Google Colab

Chat with the first page of the ACMI Public API: Google Colab.

Sample output

python chat.py
=========================
ACMI collection chat v0.1
=========================

Question: What was the collection item that has gnomes in it?
Answer: The collection item that has gnomes in it is titled "Toy magic lantern slide (Gnomes in umbrellas on water)". It is a work from Germany, circa 1900, and was last on display at ACMI: Gallery 1 on June 23, 2023. The item is categorized under the curatorial section "The Story of the Moving Image → Moving Pictures → MI-02. Play and Illusion → MI-02-C01 → Panel C8" and has measurements of 3.5 x 14.3cm. It is a 2D Object, specifically a Glass slide/Pictorial.
Sources: ['https://api.acmi.net.au/works/119591']