main <- alpha

[Anush008]

🍕 RepoQuery 🍕

A REST service to answer user-queries about public GitHub repositories

🔎 The Project

RepoQuery is an early-beta project, that uses recursive OpenAI function calling paired with semantic search using All-MiniLM-L6-V2 to index and answer user queries about public GitHub repositories.

Related Tickets & Documents

open-sauced/ai#192
open-sauced/ai#226

📬 Service Endpoints

1. /embed

To generate and store embeddings for a GitHub repository.

Parameters

The parameters are passed as a JSON object in the request body:

• owner (string, required): The owner of the repository.
• name (string, required): The name of the repository.
• branch (string, required): The name of the branch.

Response

The request is processed by the server and responses are sent as Server-sent events(SSE). The event stream will contain the following events with optional data.

repo-query/src/routes/events.rs

Lines 14 to 20 in f2f415a

	sse_events! {
	EmbedEvent,
	(FetchRepo, "FETCH_REPO"),
	(EmbedRepo, "EMBED_REPO"),
	(SaveEmbeddings, "SAVE_EMBEDDINGS"),
	(Done, "DONE"),
	}

Example

curl --location 'localhost:3000/embed' \
--header 'Content-Type: application/json' \
--data '{
    "owner": "open-sauced",
    "name": "ai",
    "branch": "beta"
}'

2. /query

To perform a query on the API with a specific question related to a repository.

Parameters

The parameters are passed as a JSON object in the request body:

• query (string, required): The question or query you want to ask.
• repository (object, required): Information about the repository for which you want to get the answer.
  • owner (string, required): The owner of the repository.
  • name (string, required): The name of the repository.
  • branch (string, required): The name of the branch.

Response

The request is processed by the server and responses are sent as Server-sent events(SSE). The event stream will contain the following events with optional data.

repo-query/src/routes/events.rs

Lines 22 to 29 in f2f415a

	sse_events! {
	QueryEvent,
	(SearchCodebase, "SEARCH_CODEBASE"),
	(SearchFile, "SEARCH_FILE"),
	(SearchPath, "SEARCH_PATH"),
	(GenerateResponse, "GENERATE_RESPONSE"),
	(Done, "DONE"),
	}

Example

curl --location 'localhost:3000/query' \
--header 'Content-Type: application/json' \
--data '{
    "query": "How is the PR description being generated using AI?",
    "repository": {
        "owner": "open-sauced",
        "name": "ai",
        "branch": "beta"
    }
}'

3. /collection

To check if a repository has been indexed.

Parameters

• owner (string, required): The owner of the repository.
• name (string, required): The name of the repository.
• branch (string, required): The name of the branch.

Response

This endpoint returns an OK status code if the repository has been indexed by the service.

Example

curl --location 'localhost:3000/embed?owner=open-sauced&name=ai&branch=beta'

🧪 Running Locally

To run the project locally, there are a few prerequisites:

• The Rust toolchain
• The Onnx Runtime. Will be downloaded and installed automatically when building the project.
• Docker to run the QdrantDB instance.

Once, the above requirements are satisfied, you can run the project like so:

Environment variables

The project requires the following environment variables to be set.

• OPENAI_API_KEY. To authenticate requests to OpenAI.

Database setup

Start Docker and run the following commands to spin-up a Docker container with a QdrantDB image.

docker pull qdrant/qdrant
docker run -p 6333:6333 -p 6334:6334 qdrant/qdrant

The database dashboard will be accessible at localhost:6333/dashboard, the project communicates with the DB on port 6334.

Running the project

Run the following command to install the dependencies and run the project on port 3000.

cargo run --release

This command will build and run the project with optimizations enabled(Highly recommended).

[jpmcb]

This is looking awesome!!

I'm unable to build due to a number of compiler errors (see nit errors left in this review)

Rust I'm building with locally:

❯ cargo --version
cargo 1.70.0 (ec8a8a0ca 2023-04-25)
❯ rustc --version
rustc 1.70.0 (90c541806 2023-05-31)

---

Looks like this is also a WIP 😅 so feel free to disregard my comments and ping me when this is ready to look at! Very cool!!

[Anush008]

Hey. I'll be updating the build steps and there's todo!(), that's breaking the build. I working on it right now. Will let you know when it's ready. Thanks for taking a look.

[jpmcb]

Quick note: we'll want to keep the bar for contributors high, so let's make sure we add a README.md and some development docs.

[Anush008]

Quick note: we'll want to keep the bar for contributors high, so let's make sure we add a README.md and some development docs.

Right. Will do.

[Anush008]

Hey @jpmcb. The project now has build steps and routes documented. Can you take a look? Once this first draft is merged, we can look into further improving the output by tweaking src/conversation/prompts.rs, moving to GPT-4, adding auth to the service using Supabase, improving the semantic search.

[jpmcb]

Awesome, this is on my docket to review today: great work!!

[jpmcb]

I'm not seeing an attribution or license on model/model_quantized.onnx in https://huggingface.co/rawsh/multi-qa-MiniLM-distill-onnx-L6-cos-v1

Is this ok to use in our project? Are there any constraints on us deploying it?

---

Overall, this looks really great. Well done. I'm planning to do another, more in depth pass this afternoon but wanted to drop this early feedback for you.

[Anush008]

I'm not seeing an attribution or license on model/model_quantized.onnx in https://huggingface.co/rawsh/multi-qa-MiniLM-distill-onnx-L6-cos-v1

Is this ok to use in our project? Are there any constraints on us deploying it?

https://www.sbert.net/#citing-authors mentions

If you use one of the multilingual models, feel free to cite our publication Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation:

I've updated the README.md with the attributions. 02db5c6.

[jpmcb]

Overall, looks good. Once existing comments are addressed, will merge 👍🏼