Skip to content

How to train on other documentation

Anton Larin edited this page Jun 14, 2023 · 21 revisions

How to train on other documentation

This AI can use any documentation, but first it needs to be prepared for similarity search.

video-example-of-how-to-do-it

Start by going to /scripts/ folder

If you open this file you will see that it uses RST files from the folder to create a index.faiss and index.pkl.

It currently uses OPEN_AI to create vector store, so make sure your documentation is not too big. Pandas cost me around 3-4$

You can usually find documentation on github in docs/ folder for most open-source projects.

1. Find documentation in .rst/.md and create a folder with it in your scripts directory

Name it inputs/
Put all your .rst/.md files in there
The search is recursive, so you don't need to flatten them

If there are no .rst/.md files just convert whatever you find to txt and feed it. (don't forget to change the extension in script)

2. Create .env file in scripts/ folder

And write your OpenAI API key inside OPENAI_API_KEY=<your-api-key>

3. Run scripts/ingest.py

python ingest.py ingest

It will tell you how much it will cost

4. Move index.faiss and index.pkl generated in scripts/output to application/ folder.

5. Run web app

Once you run it will use new context that is relevant to your documentation Make sure you select default in the dropdown in the UI

Customisation

You can learn more about options while running ingest.py by running:

python ingest.py --help

Options
ingest Runs 'ingest' function converting documentation to to Faiss plus Index format
--dir TEXT List of paths to directory for index creation. E.g. --dir inputs --dir inputs2 [default: inputs]
--file TEXT File paths to use (Optional; overrides directory) E.g. --files inputs/1.md --files inputs/2.md
--recursive / --no-recursive Whether to recursively search in subdirectories [default: recursive]
--limit INTEGER Maximum number of files to read
--formats TEXT List of required extensions (list with .) Currently supported: .rst, .md, .pdf, .docx, .csv, .epub, .html [default: .rst, .md]
--exclude / --no-exclude Whether to exclude hidden files (dotfiles) [default: exclude]
-y, --yes Whether to skip price confirmation
--sample / --no-sample Whether to output sample of the first 5 split documents. [default: no-sample]
--token-check / --no-token-check Whether to group small documents and split large. Improves semantics. [default: token-check]
--min_tokens INTEGER Minimum number of tokens to not group. [default: 150]
--max_tokens INTEGER Maximum number of tokens to not split. [default: 2000]
convert Creates documentation in .md format from source code
--dir TEXT Path to a directory with source code. E.g. --dir inputs [default: inputs]
--formats TEXT Source code language from which to create documentation. Supports py, js and java. E.g. --formats py [default: py]