mettis

mettisDB: Distributed lite vector database built from *scratch.

What it does:

• full-text search using proximity ranking
• semantic search via HNSW + Cosine distance
• integrated basic text embedding service (Python HTTP API around a sentence transformer)
• Reciprocal Rank Fusion for merging full-text + semantic search results
• in-memory serving + disk persistence
• fault-tolerance with segment replication using Raft

What it's not:

• production-ready (code is pretty sus right now)
• perfect (because enemy of good)

Architecture

Getting started

Install the dependencies for the basic text embedding service in third_party folder using pip.

pip install -r requirements.txt

Then start the service like so:

uvicorn main:app

Set an environment variable EmbeddingHost which points to the address of the embedding service

export EmbeddingHost="http://127.0.0.1:8000/embeddings"

Proceed to start instance(s) of the vector db

flags

• httpAddr: address of HTTP API service
• joinAddr: HTTP API service address of primary node to join
• nodeId: unique identifier for node
• raftAddr: raft address for node

Run single-node

go run cmd/server/main.go -httpAddr 127.0.0.1:8111 -nodeId 0 -raftAddr 127.0.0.1:9000

API

POST /index

index a document

curl --location '127.0.0.1:8111/index' --header 'Content-Type: application/json' --data '{"text": "some text"}'

GET /search

do a search

curl --location --request GET '127.0.0.1:8111/search' 
--header 'Content-Type: application/json' 
--data '{"query": "some text"}'

Run 3-node cluster

Run the commands below on different machines (at least different instances of the project to simulate)

go run cmd/server/main.go -httpAddr 127.0.0.1:8111 -nodeId 0 -raftAddr 127.0.0.1:9000

Replicas join the primary on 127.0.0.1:9000

go run cmd/server/main.go -httpAddr 127.0.0.1:8112 -nodeId 1 -raftAddr 127.0.0.1:9001 -joinAddr 127.0.0.1:8111

go run cmd/server/main.go -httpAddr 127.0.0.1:8113 -nodeId 2 -raftAddr 127.0.0.1:9002 -joinAddr 127.0.0.1:8111

TODO

• Indexing
  • Concurrent indexing using goroutines to process terms
• Retrieval
  • Boolean queries
  • Concurrent memtable search
• Ranking
• API
  • Bulk index
  • Document deletion
• Storage
  • Segment compaction
• Replication
  • Snapshot working?
• Deployment
  • Containerisation
• Code quality
  • Penance for all the atrocities I committed.

*Would not have been possible without these resources:

• Let's build a Full-Text Search engine
• Exploring Memtables
• hraftd
• Hierarchical Navigable Small Worlds
• Information Retrieval Implementing and Evaluating Search Engines
• Distributed Services with Go