Turn PostgreSQL into your search engine in a Pythonic way.

Installation

pip install vechord

Features

• vector search with RaBitQ (powered by VectorChord)
• multivec search with WARP (powered by VectorChord)
• keyword search with BM25 score (powered by VectorChord-bm25)
• reduce boilerplate code by taking full advantage of the Python type hint
• provide decorator to inject the data from/to the database
• guarantee the data consistency with transaction
• auto-generate the web service
• provide common functions like (can also use any other libraries as you like):
  • Augmenter for contextual retrieval
  • Chunker to segment the text into chunks
  • Embedding to generate the embedding from the text
  • Evaluator to evaluate the search results with NDCG, MAP, Recall, etc.
  • Extractor to extract the content from PDF, HTML, etc.
  • Reranker for hybrid search

Examples

• simple.py: for people that are familiar with specialized vector database APIs
• beir.py: the most flexible way to use the library (loading, indexing, querying and evaluation)
• web.py: build a web application with from the defined tables and pipeline
• essay.py: extract the content from Paul Graham's essays and evaluate the search results from LLM generated queries
• contextual.py: contextual retrieval example
• hybrid.py: hybrid search that rerank the results from vector search with keyword search

User Guide

For the API references, check our documentation.

Define the table

from typing import Annotated, Optional
from vechord.spec import Table, Vector, PrimaryKeyAutoIncrease, ForeignKey

# use 768 dimension vector
DenseVector = Vector[768]

class Document(Table, kw_only=True):
    uid: Optional[PrimaryKeyAutoIncrease] = None  # auto-increase id, no need to set
    link: str = ""
    text: str

class Chunk(Table, kw_only=True)
    uid: Optional[PrimaryKeyAutoIncrease] = None
    doc_id: Annotated[int, ForeignKey[Document.uid]]  # reference to `Document.uid`
    vector: DenseVector  # this comes with a default vector index
    text: str

Inject with decorator

import httpx
from vechord.registry import VechordRegistry
from vechord.extract import SimpleExtractor
from vechord.embedding import GeminiDenseEmbedding

vr = VechordRegistry(namespace="test", url="postgresql://postgres:[email protected]:5432/")
# ensure the table and index are created if not exists
vr.register([Document, Chunk])
extractor = SimpleExtractor()
emb = GeminiDenseEmbedding()

@vr.inject(output=Document)  # dump to the `Document` table
# function parameters are free to define since `inject(input=...)` is not set
def add_document(url: str) -> Document:  # the return type is `Document`
    with httpx.Client() as client:
        resp = client.get(url)
        text = extractor.extract_html(resp.text)
        return Document(link=url, text=text)

@vr.inject(input=Document, output=Chunk)  # load from the `Document` table and dump to the `Chunk` table
# function parameters are the attributes of the `Document` table, only defined attributes
# will be loaded from the `Document` table
def add_chunk(uid: int, text: str) -> list[Chunk]:  # the return type is `list[Chunk]`
    chunks = text.split("\n")
    return [Chunk(doc_id=uid, vector=emb.vectorize_chunk(t), text=t) for t in chunks]

if __name__ == "__main__":
    add_document("https://paulgraham.com/best.html")  # add arguments as usual
    add_chunk()  # omit the arguments since the `input` is will be loaded from the `Document` table
    vr.insert(Document(text="hello world"))  # insert manually
    print(vr.select_by(Document.partial_init()))  # select all the columns from table `Document`

Transaction

To guarantee the data consistency, users can use the VechordRegistry.run method to run multiple functions in a transaction.

In this transaction, all the functions will only load the data from the database that is inserted in the current transaction. So users can focus on the data processing part without worrying about which part of data has not been processed yet.

vr.set_pipeline([add_document, add_chunk])
vr.run("https://paulgraham.com/best.html")  # only accept the arguments for the first function

Search

print(vr.search_by_vector(Chunk, emb.vectorize_query("startup")))

Customized Index Configuration

from vechord.spec import VectorIndex

class Chunk(Table, kw_only=True):
    uid: Optional[PrimaryKeyAutoIncrease] = None
    vector: Annotated[DenseVector, VectorIndex(distance="cos", lists=128)]
    text: str

HTTP Service

This creates a WSGI application that can be served by any WSGI server.

Open the OpenAPI Endpoint to check the API documentation.

from vechord.service import create_web_app
from wsgiref.simple_server import make_server

app = create_web_app(vr)
with make_server("", 8000, app) as server:
    server.serve_forever()

Development

docker run --rm -d --name vdb -e POSTGRES_PASSWORD=postgres -p 5432:5432 ghcr.io/tensorchord/vchord_bm25-postgres:pg17-v0.1.1
envd up
# inside the envd env, sync all the dependencies
make sync
# format the code
make format