Cross-Modal Vector Fusion: Joint Embeddings for Image and Text Retrieval

Unifying Text and Image Semantic Search

In legacy search architectures:

• Separated Data Silos: Image metadata relies on alt-text tags, while text documents use vector embeddings, preventing cross-media semantic queries.
• Mismatched Scoring: Combining text similarity scores with image classification confidence yields inconsistent ranking outputs.
• The Solution: Shared Embedding Spaces. Joint multi-modal encoders project image features and text tokens into the same high-dimensional coordinate space. A query like "database schema diagram" matches image vectors of architecture diagrams without requiring text OCR.

1. Projecting into Shared Vector Spaces

To implement cross-modal retrieval:

• Normalize Vector Projections: Ensure text and image embeddings are normalized to unit length so dot-product calculations equal cosine similarity.
• Unified Vector Indices: Store image vectors and text vectors inside a single HNSW (Hierarchical Navigable Small World) index, tagged with media type metadata.

2. Executing Cross-Modal Queries

The multi-modal search manager coordinates retrieval:

1. Embed User Query: Pass text inputs to the joint text encoder to generate a 512-dimensional query vector.
2. Retrieve Mixed Assets: Perform a single vector search against the unified index to return matching images and text chunks simultaneously.

Code Demo: Cross-Modal Vector Search Engine

Below is a Python implementation of a cross-modal search engine. It projects simulated text and image vectors into a shared index, calculates cosine similarity, and retrieves multi-modal assets.

import numpy as np
from typing import List, Dict, Any

class CrossModalVectorSearch:
    def __init__(self, vector_dim: int = 512):
        self.vector_dim = vector_dim
        # Unified vector index storing both text and image assets
        self.index: List[Dict[str, Any]] = []

    def _normalize_vector(self, vec: np.ndarray) -> np.ndarray:
        norm = np.linalg.norm(vec)
        return vec / norm if norm > 0 else vec

    def add_asset(self, asset_id: str, media_type: str, raw_vector: np.ndarray, metadata: Dict[str, Any]):
        normalized_vec = self._normalize_vector(raw_vector)
        self.index.append({
            "id": asset_id,
            "media_type": media_type,
            "vector": normalized_vec,
            "metadata": metadata
        })
        print(f"📦 [Index] Added {media_type.upper()} asset '{asset_id}' to unified vector space.")

    def search_by_text(self, text_query_vector: np.ndarray, top_k: int = 2) -> List[Dict[str, Any]]:
        query_vec = self._normalize_vector(text_query_vector)
        results = []

        # Perform cosine similarity search across shared vector space
        for item in self.index:
            similarity = np.dot(query_vec, item["vector"])
            results.append({
                "id": item["id"],
                "media_type": item["media_type"],
                "score": float(similarity),
                "metadata": item["metadata"]
            })

        # Sort results by similarity score
        results.sort(key=lambda x: x["score"], reverse=True)
        return results[:top_k]

if __name__ == "__main__":
    search_engine = CrossModalVectorSearch(vector_dim=4)

    # 1. Populate unified index with image and text vectors
    # Image asset vector (e.g. system architecture diagram screenshot)
    image_vector = np.array([0.9, 0.1, 0.0, 0.2])
    search_engine.add_asset(
        asset_id="img_arch_diag",
        media_type="image",
        raw_vector=image_vector,
        metadata={"filename": "architecture_diagram.png", "resolution": "1920x1080"}
    )

    # Text asset vector (e.g. documentation paragraph)
    text_vector = np.array([0.1, 0.8, 0.2, 0.1])
    search_engine.add_asset(
        asset_id="doc_setup_guide",
        media_type="text",
        raw_vector=text_vector,
        metadata={"title": "Local Deployment Setup Guide"}
    )

    # 2. Execute text query vector (closely matching the image vector)
    query_text_vector = np.array([0.85, 0.15, 0.05, 0.1])

    print("\n🛡️ Executing Cross-Modal Vector Search...")
    print("------------------------------------------")

    matches = search_engine.search_by_text(query_text_vector, top_k=2)

    print("\n📈 --- Unified Multi-Modal Search Results ---")
    for idx, match in enumerate(matches):
        print(f"    [Match {idx + 1}] ID: {match['id']} ({match['media_type'].upper()}) | Similarity Score: {match['score']:.4f}")
        print(f"               Metadata: {match['metadata']}")

Multi-Modal Vector Takeaways

• Project to Shared Spaces: Use joint embedding encoders (CLIP/ImageBind) to map text and images into the same coordinate space.
• Normalize Vector Outputs: Normalize embeddings to unit length before indexing to compute cosine similarity using fast dot products.
• Tag Media Metadata: Store media type tags alongside vector records to enable filtering by asset type.