Multimodel fusion

Combine search results from multiple embedding models using fusion.

Multimodel fusion runs the same query through multiple embedding models and fuses the results into a single ranking. Different embedding models capture different aspects of semantic meaning, and combining their results produces more comprehensive retrieval than any single model alone.

This code example creates a collection, inserts 100 sample documents, and generates three separate query embeddings to simulate different models. It searches with each embedding, then fuses all three result sets using RRF to produce a single combined ranking.

In production, replace the simulated embeddings with actual embedding model outputs. The example below uses random vectors as placeholders for models like OpenAI text-embedding-3-small, Cohere embed-multilingual-v3.0, and sentence-transformers/all-MiniLM-L6-v2.

Python

import asyncio
import random
from actian_vectorai import AsyncVectorAIClient, VectorParams, Distance, PointStruct, reciprocal_rank_fusion

COLLECTION = "articles"
DIMENSION = 128

async def main():
    async with AsyncVectorAIClient("localhost:6574") as client:
        # Create collection if it doesn't exist
        if not await client.collections.exists(COLLECTION):
            await client.collections.create(
                COLLECTION,
                vectors_config=VectorParams(size=DIMENSION, distance=Distance.Cosine)
            )

            # Insert sample points
            points = [
                PointStruct(
                    id=i,
                    vector=[random.gauss(0, 1) for _ in range(DIMENSION)],
                    payload={"text": f"Article {i}"}
                )
                for i in range(1, 101)
            ]
            await client.points.upsert(COLLECTION, points)
            print(f"✓ Inserted {len(points)} points")

        # Simulate embeddings from different models
        # In practice, use actual embedding models like:
        # - OpenAI text-embedding-3-small
        # - Cohere embed-multilingual-v3.0
        # - sentence-transformers/all-MiniLM-L6-v2

        openai_embedding = [random.gauss(0, 1) for _ in range(DIMENSION)]
        cohere_embedding = [random.gauss(0.2, 0.9) for _ in range(DIMENSION)]
        sentence_transformer_embedding = [random.gauss(-0.1, 1.1) for _ in range(DIMENSION)]

        # Search with each model
        results = []
        for embedding in [openai_embedding, cohere_embedding, sentence_transformer_embedding]:
            result = await client.points.search(
                COLLECTION,
                vector=embedding,
                limit=15
            )
            results.append(result)

        # Fuse all results
        final_results = reciprocal_rank_fusion(results, ranking_constant_k=60, limit=10)

        print(f"Combined results from 3 embedding models: {len(final_results)} unique documents")
        for i, point in enumerate(final_results[:5], 1):
            print(f"{i}. ID: {point.id}, Score: {point.score:.4f}")

asyncio.run(main())

Each fused result includes these fields:

• id: The unique identifier of the matching point
• score: Fused score combining rank positions from all three model searches
• payload: Metadata object from the matching point

Multimodel fusion provides these advantages:

• Different models capture complementary semantic signals
• Results that rank highly across multiple models are more likely to be relevant
• Reduces the risk of missing relevant documents that one model ranks poorly