Skip to content

Verifiable nearest-neighbor search with Aligned Layer & SP1.

License

Notifications You must be signed in to change notification settings

erhant/aligned-vnns

 
 

Repository files navigation

Verifiable Nearest Neighbor Search

A verifiable nearest-neighbor search program, powered by SP1.

This project was built for Aligned Builders Hackathon:

The requested information as per the judging criteria has been given within the project link above.

Setup

Installation

You need the following:

Keystore

You can create a new wallet & store it directly in a random file with:

PRIV_KEY=$(cast wallet new | grep "Private key" | cut -d ' ' -f 3)

# will prompt for a password, can be left empty
cast wallet import --private-key $PRIV_KEY -k ./secrets/ $RANDOM.json

It will prompt for a password, and then output the keystore under secrets folder.

You can view the respective address of a keystore file with:

# will prompt for the password
cast wallet address --keystore ./secrets/wallet.json

Aligned SDK

First, deposit some funds to Aligned layer (skip this if you already have done so):

./aligned deposit ./path/to/keystore.json

This will print a transaction hash, which can be viewed at https://holesky.etherscan.io/tx/<hash-here>.

You can check your balance with:

./aligned balance ./path/to/keystore.json

Embeddings

Warning

You need to have Ollama running on localhost:11434 (default) to run the commands in this section.

The repository comes with existing embeddings under the data folder, within the files with .index.json extension. For this project, each data has the following type:

{
  name: string;
  description: string;
}

You can create your own embeddings as follows:

cargo run --bin vnns-embedder index -p ./path/to/data.json
# will output ./path/to/data.index.json

Generate Query Vector

To generate a query vector to be used within a proof, use the following command:

cargo run --bin vnns-embedder query -p ./path/to/data.json -t "your text to be converted here"
# will output ./path/to/data.query.json

This saves the vector itself within the JSON file, which the prover reads from disk.

Usage

Build

To build the VNNS program, run the following command:

cd program
cargo prove build --elf-name riscv32im-succinct-vnns-elf

To build the aggregator program:

cd aggregator
cargo prove build --elf-name riscv32im-succinct-aggregator-elf

Execute

To run the program without generating a proof:

RUST_LOG=info cargo run --bin vnns-script --release -- --execute --path ./data/foods-small.json

This will execute the program and display the output.

Prove

To generate a core proof for your program:

RUST_LOG=info cargo run --bin vnns-script --release -- --prove --path ./data/foods-small.json

This will generate many proofs (based on file size & batch size) and store them under the same directory as given in path. To see which text the result belongs to, copy the Output Commitment on the console, and look-up the item within the vector index that has the same hash with that commitment.

Tip

You can configure the batch size with --batch-size <number> argument, default is 4. The batch size should be small especially if the vector is large (1000s of elements) because they are all of type f32 and will consume a lot of resources within the zkVM.

Tip

If --aggregate option is passed, it will aggregate and store the final proof as well with the extension .agg.proof and .agg.pub.

Submit

Consider proofs generated for some data ./data.json. You can submit all batches of proofs to Aligned Layer with:

./aligned.sh submit ./path/to/keystore.json ./data.json

To send the aggregated proof only, you can use:

./aligned.sh submit-agg ./path/to/keystore.json ./data.json

Note

For each proof, it will ask for your keystore password.

License

The project is MIT licensed as per SP1 project template.

About

Verifiable nearest-neighbor search with Aligned Layer & SP1.

Topics

Resources

License

Stars

Watchers

Forks