pipx install git+https://github.com/Storia-AI/sage.git@main

python -m venv sage-venv
source sage-venv/bin/activate
git clone https://github.com/Storia-AI/sage.git
cd sage
pip install -e .

1. To index the codebase locally, we use the open-source project Marqo, which is both an embedder and a vector store. To bring up a Marqo instance:

   docker rm -f marqo
   docker pull marqoai/marqo:latest
   docker run --name marqo -it -p 8882:8882 marqoai/marqo:latest
   

   This will open a persistent Marqo console window. This should take around 2-3 minutes on a fresh install.

2. To chat with an LLM locally, we use Ollama:

   • Head over to ollama.com to download the appropriate binary for your machine.
   • Open a new terminal window
   • Pull the desired model, e.g. ollama pull llama3.1.

1. For embeddings, we support OpenAI and Voyage. According to our experiments, OpenAI is better quality. Their batch API is also faster, with more generous rate limits. Export the API key of the desired provider:

   export OPENAI_API_KEY=... # or
   export VOYAGE_API_KEY=...
   

2. We use Pinecone for the vector store, so you will need an API key:

   export PINECONE_API_KEY=...
   

   If you want to reuse an existing Pinecone index, specify it. Otherwise we'll create a new one called sage.

   export PINECONE_INDEX_NAME=...
   

3. For reranking, we support NVIDIA, Voyage, Cohere, and Jina.

   • According to our experiments, NVIDIA performs best. To get an API key, follow these instructions. Note that NVIDIA's API keys are model-specific. We recommend using nvidia/nv-rerankqa-mistral-4b-v3.
   • Export the API key of the desired provider:

   export NVIDIA_API_KEY=...  # or
   export VOYAGE_API_KEY=...  # or
   export COHERE_API_KEY=...  # or
   export JINA_API_KEY=...
   

4. For chatting with an LLM, we support OpenAI and Anthropic. For the latter, set an additional API key:

   export ANTHROPIC_API_KEY=...
   

For easier configuration, adapt the entries within the sample .sage-env (change the API keys names based on your desired setup) and run:

source .sage-env

To index and chat with a private repository, simply set the GITHUB_TOKEN environment variable. To obtain this token, go to github.com > click on your profile icon > Settings > Developer settings > Personal access tokens. You can either make a fine-grained token for the desired repository, or a classic token.

export GITHUB_TOKEN=...

You can specify an inclusion or exclusion file in the following format:

# This is a comment
ext:.my-ext-1
ext:.my-ext-2
ext:.my-ext-3
dir:my-dir-1
dir:my-dir-2
dir:my-dir-3
file:my-file-1.md
file:my-file-2.py
file:my-file-3.cpp

where:

• ext specifies a file extension
• dir specifies a directory. This is not a full path. For instance, if you specify dir:tests in an exclusion directory, then a file like /path/to/my/tests/file.py will be ignored.
• file specifies a file name. This is also not a full path. For instance, if you specify file:__init__.py, then a file like /path/to/my/__init__.py will be ignored.

To specify an inclusion file (i.e. only index the specified files):

sage-index $GITHUB_REPO --include=/path/to/inclusion/file

To specify an exclusion file (i.e. index all files, except for the ones specified):

sage-index $GITHUB_REPO --exclude=/path/to/exclusion/file

By default, we use the exclusion file sample-exclude.txt.

You will need a GitHub token first:

export GITHUB_TOKEN=...

To index GitHub issues without comments:

sage-index $GITHUB_REPO --index-issues

To index GitHub issues with comments:

sage-index $GITHUB_REPO --index-issues --index-issue-comments

To index GitHub issues, but not the codebase:

sage-index $GITHUB_REPO --index-issues --no-index-repo

Retrieving the right files from the vector database is arguably the quality bottleneck of the system. We are actively experimenting with various retrieval strategies and documenting our findings here.

Currently, we support the following types of retrieval:

• Vanilla RAG from a vector database (nearest neighbor between dense embeddings). This is the default.

• Hybrid RAG that combines dense retrieval (embeddings-based) with sparse retrieval (BM25). Use --retrieval-alpha to weigh the two strategies.

  • A value of 1 means dense-only retrieval and 0 means BM25-only retrieval.
  • Note this is not available when running locally, only when using Pinecone as a vector store.
  • Contrary to Anthropic's findings, we find that BM25 is actually damaging performance on codebases, because it gives undeserved advantage to Markdown files.

• Multi-query retrieval performs multiple query rewrites, makes a separate retrieval call for each, and takes the union of the retrieved documents. You can activate it by passing --multi-query-retrieval. This can be combined with both vanilla and hybrid RAG.

  • We find that on our benchmark this only marginally improves retrieval quality (from 0.44 to 0.46 R-precision) while being significantly slower and more expensive due to LLM calls. But your mileage may vary.

• LLM-only retrieval completely circumvents indexing the codebase. We simply enumerate all file paths and pass them to an LLM together with the user query. We ask the LLM which files are likely to be relevant for the user query, solely based on their filenames. You can activate it by passing --llm-retriever.

  • We find that on our benchmark the performance is comparable with vector database solutions (R-precision is 0.44 for both). This is quite remarkable, since we've saved so much effort by not indexing the codebase. However, we are reluctant to claim that these findings generalize, for the following reasons:
    • Our (artificial) dataset occasionally contains explicit path names in the query, making it trivial for the LLM. Sample query: "Alice is managing a series of machine learning experiments. Please explain in detail how main in examples/pytorch/image-pretraining/run_mim.py allows her to organize the outputs of each experiment in separate directories."
    • Our benchmark focuses on the Transformers library, which is well-maintained and the file paths are often meaningful. This might not be the case for all codebases.