buster / README.md
hbertrand's picture
mv
97a0e6c
|
raw
history blame
1.78 kB
# Buster, the Mila cluster chatbot!
Buster is a chatbot that can answer questions about the Mila cluster. You can try it [here](https://huggingface.co/spaces/jerpint/buster).
![Question about tmpdir](buster/imgs/qa_tmpdir.png)
<!---
Where should I put my data when running a job?
-->
![Question about GPUs jobs](buster/imgs/qa_gpus.png)
<!---
How do I run a job with 2 GPUs?
-->
It also works in french, although not as well:
![Question en français](buster/imgs/qa_french.png)
<!---
Comment est-ce que je lance une job avec 2 GPUs?
-->
If you ask a stupid question, you will get a stupid answer:
![Dumb question](buster/imgs/qa_pasta.png)
<!---
What has more calories: a DGX A100 or a pasta salad?
-->
## How does Buster works?
First, we parsed the [Mila docs](https://docs.mila.quebec/index.html) into snippets. You can see exactly how in [`buster.docparser.get_all_documents`](https://github.com/jerpint/buster/blob/main/buster/docparser.py#L17).
The resulting snippets can be seen in [`buster/data/documents.csv`](https://github.com/jerpint/buster/blob/main/buster/data/documents.csv).
For each snippet, we obtain an embedding by using the [OpenAI API](https://beta.openai.com/docs/guides/embeddings/what-are-embeddings).
Then, when a user asks a question, we compute its embedding, find the snippet from the doc with the highest cosine similarity to the question.
Finally, we craft the prompt:
- The most relevant snippet from the doc
- The engineering prompt: "Now answer the following question:"
- The user's question
We send the prompt to the [OpenAI API](https://beta.openai.com/docs/api-reference/completions), and display the answer to the user!
### Currently used models
- For embeddings: "text-embedding-ada-002"
- For completion: "text-davinci-003"