A good starting point for building your own personal AI.

Unless you ask it to search Wikipedia, etc., no internet connection is required.

Why is this important? Because as long as a chatbot still sends information to the cloud (OpenAI, Azure, ...), I wouldn't trust it with sensitive info like my paystubs, health records, passwords, etc.

Uses off-the-shelf components to keep the codebase small and easy to understand.

• LM Studio serves a LLM (Zephyr 7B beta, as of Jan 2024) locally. No privacy concerns there -- The LLM is not fine-tuned with any private information, and the server is stateless. (Note: You can easily replace LM Studio with Ollama, etc., but I like the GUI that LM Studio provides.)
• LlamaIndex provides a natural-language querying interface. It also indexes documents into embeddings, which are stored into a Chroma database.
• ChainLit provides a web UI to LlamaIndex that looks like ChatGPT. (I tried building a TUI with rich, but it's not very obvious how to surface the retrieved documents.)

Why is this important? Because you almost certainly have your own niche needs for your own AI chatbot, so you are most likely to be developing something solely on your own. With limited workforce, it's important to keep the codebase small and easy to understand.

Ask it:

Name a type of drink that I enjoy, and then look up its country of origin. Be concise.

and it will say:

Based on the available evidence, it appears that coffee may have originated in either Ethiopia or Yemen. However, as the legend of Kaldi suggests, Ethiopia has long been associated with coffee's history.

after looking up the user's personal notes and then consulting Wikipedia.

Ensure that you have an OpenAI-compatible inference server running at http://localhost:1234/. If you're using LM Studio, a tried-and-true configuration looks like this:

Tips: The LLM may be biased to generate "Observations" when it is only supposed to generate up to "Action Input". To mitigate this problem, add "Observation:" as a Stop String in LM Studio, so that LM Studio will stop the LLM from generating any more text after it sees "Observation:".

Then, run the script:

chainlit run main.py -w

This repo uses pre-commit hooks to automate many chores. Install them.

As a Python-based project, this repo registers all its dependencies in the pyproject.toml file. Use Poetry to install them.

PYTHONPATH=. poetry install --no-root

We use --no-root because we don't want to install the project itself as a dependency. It's an application, not a library.

As this article explains:

The main use of PYTHONPATH is when we are developing some code that we want to be able to import from Python, but that we have not yet made into an installable Python package.

main.py is the entrypoint. It runs a ReAct Agent (S Yao, et, al.). As an agent, the AI is capable of wielding several tools. For example,

• It can use tool_for_my_notes.py to look up plain-text notes you stored in a folder. For demo purposes, the folder demo-notes/ contains some stubs that you can check out.
• It can use tool_for_wikipedia.py to find answers to a given question after consulting Wikipedia articles.

chainlit.md and public/ are simply UI assets for the web frontend.

To improve the precision of tool_for_my_notes.py, I modified the default prompt for the sub-question query engine in LlamaIndex by asking it to generate keywords rather than complete sentences. The changes are in sub_question_generating_prompt_in_keywords.py.

Similarly, I also overrode the agent-level system prompt. Since it's quite a long prose, I put that in a separate file, system_prompt.md.

One-shot examples, instead of zero-shot. In both the QueryEngineTool in tool_for_my_notes.py and the OnDemandLoaderTool in tool_for_wikipedia.py, I added one example in the tool description. This greatly improves the quality of Action Inputs generated by the LLM.

Generated via cloc --md .: