Skills/chroma/SKILL.md

name, description

name	description
chroma	This skill provides access to a RAG-like context enhancer that uses Chroma locally.

Chroma

Whenever the user asks to "use chroma", you should perform a search on Chroma, which runs locally via a Python script. The working directory you must use is located in ~/code/chroma and the entrypoint is main.py. The script MUST be invoked using uv.

You have access to these commands:

• $ uv run main.py lc -> Lists the existing collections.
• $ uv run main.py q <collection> <query> -> Performs a query. Be sure to quote the <query> if this is composed by multiple words.

Then use the response from Chroma to enhance the context and give the user a refined response.

A note on file sources

The Chroma response returns the metadatas for the chunks it finds. Among these metadatas, there is file_name, which refers to the original file that was chunked and imported. DO NOT ATTEMPT to find or fetch these files. They most likely do not exist in the filesystem. You SHOULD ALWAYS however cite correctly from which files (ONLY from Chroma's metadatas) the information is coming.

Example use case

START

User query:

Search in chroma information about lovecraft's Dunwich horror.

Step 1: Get the available collections with uv run main.py lc. The output is:

lovecraft
documents

Most likely our information is in the lovecraft collection. We will use that for the query.

Step 2: Query using uv run main.py q lovecraft <query>. The query is up to you, create one keeping into account that this is a raw query on a vector DB. Be concise, extract keywords, avoid noise.

Step 3: Get the results, enhance the context, and respond to the user.

END