Need help learning how to work with embeddings

[rpdelaney]

I don't really understand what I'm doing here:

$ llm embed -i bash_manual.txt -m text-embedding-ada-002 --format blob --store ~/.config/io.datasette.llm/embeddings.db
Error: Must provide both collection and id
$

My practical goal is to use llm for local RAG operations against manuals and technical documentation. My overall goal is to learn how to work with these tools :) Any help is appreciated!

[Muguett-DBY]

The confusing part is --store: it's a flag meaning “also save the original text”, not the database path. In your command the path is being read as a collection name, so llm embed asks for an ID as well. Use -d for the database path, and provide a collection and an ID:

llm embed manuals bash-excerpt -i manual-excerpt.txt \
  -m text-embedding-ada-002 --store -d embeddings.db

Here manuals is the collection and bash-excerpt is the ID. --format blob controls output to the terminal; you don't need it to save the vector. Start with a short excerpt: an entire manual may exceed the embedding model's input limit.

For retrieval over a manual, split it into smaller passages and give each a stable ID, for example chunks.csv:

id,content
bash-pipes-1,Text about pipes goes here
bash-redirection-1,Text about redirection goes here

Then index and search those passages:

llm embed-multi manuals chunks.csv --format csv \
  -m text-embedding-ada-002 --store -d embeddings.db
llm similar manuals -c "How do pipes work?" -n 3 -d embeddings.db

The results contain matching IDs, scores and the saved text. You can pass the relevant passages to a separate language-model prompt as context; similarity search alone doesn't generate an answer. A collection keeps the embedding model it was created with, so use a new collection if you switch models.

One distinction for “local RAG”: SQLite storage is local, but text-embedding-ada-002 sends text to OpenAI and needs an API key. For local embeddings, the sentence-transformers plugin is one option; it downloads a model and dependencies to run locally. See the embedding CLI guide for setup and model choices.

I reproduced the original error and checked the corrected single-item, CSV and search commands with LLM 0.36 and a small offline test model, using Codex assistance. I didn't test OpenAI or assess retrieval quality.