Skip to content

Latest commit

 

History

History
112 lines (74 loc) · 3.88 KB

README.md

File metadata and controls

112 lines (74 loc) · 3.88 KB
Raw

Minimal Local RAG using llamafile

A very basic interactive CLI for indexing and querying documents using llamafiles for embeddings and text generation. Index is based on a FAISS vector store. Default embedding model is mxbai-embed-large-v1 (llamafile link) and text generation model is mistral-7b-instruct-v0.2 (llamafile link). (These can be changed by editing setup.sh.)

Setup:

cp .env.example .env
./setup.sh

This script will download llamafiles from HuggingFace and may take several minutes depending on your internet connection.

NOTE: setup script requires pyenv

Quickstart with toy data

To start the app, run:

./app.sh

When you run the app, it will:

  1. Start two llamafile servers on separate ports, one for the embedding model (port 8080) and one for the text generation model (port 8081). This might take ~40 seconds.
  2. If it's the first time you're running the app, it will automatically ingest the contents of the files in the toy_data/ directory into a vector store (the "index"). Contents of the toy_data/ directory:
1.txt: Alice likes red squares.
2.txt: Bob likes blue circles.
3.txt: Chris likes blue triangles.
4.txt: David does not like green triangles.
5.txt: Mary does not like circles.
  1. After that's done, it will start an interactive CLI that allows you to ask a model questions about the data in the index. The CLI should look like:
Enter query (ctrl-d to quit): [What does Alice like?]>

If you just hit Enter here, by default the query will be "What does Alice like?". The app output should look like:

=== Query ===
What does Alice like?

=== Search Results ===
0.7104 - " alice likes red squares ."
0.5229 - " bob likes blue circles ."
0.4088 - " chris likes blue triangles ."

=== Prompt ===
"You are an expert Q&A system. Answer the user's query using the provided context information.
Context information:
 alice likes red squares .
 bob likes blue circles .
 chris likes blue triangles .
Query: What does Alice like?"
(prompt_ntokens: 55)


=== Answer ===
"
Answer: Alice likes red squares."

--------------------------------------------------------------------------------

Here some other queries you could try:

  • Who hates three-sided shapes?
  • Who likes shapes that are the color of the sky?
  • Who likes rectangles?

That's pretty much it.

App Configuration

You can change most app settings via the .env file. The default file should look like:

EMBEDDING_MODEL_PORT=8080
GENERATION_MODEL_PORT=8081
INDEX_LOCAL_DATA_DIRS=local_data,toy_data
INDEX_TEXT_CHUNK_LEN=128
INDEX_SAVE_DIR=./index-toy

See settings.py for all available options.

Using different models

By default, the app uses:

  • Embeddings:
  • Text generation:

Adding your own data

By default, the app is configured to index the contents of the directories listed in INDEX_LOCAL_DATA_DIRS, which are local_data and toy_data. Currently we only support indexing .txt files.

First, in your .env, change INDEX_SAVE_DIR to wherever you want your index to be saved. The app will not change or overwrite an existing index, so either change the directory in the .env or delete the existing index at ./index-toy.

There are 2 ways to add data:

  1. Add .txt files to the local_data/ directory. You can remove toy_data/ from the INDEX_LOCAL_DATA_DIRS list in our .env file. You can also just add another directory to the INDEX_LOCAL_DATA_DIRS list.
  2. Add web pages to the index by specifying one or more URLs to the INDEX_URLS var in your .env file, e.g. INDEX_URLS=url1,url2,....