Vector Retriever - Milvus Microservice#

GitHub

Retrieve relevant visual data from a vector database using text or image queries

Overview#

The Vector Retriever microservice is designed to search visual data efficiently by querying embeddings stored in a vector database. It uses the CLIP model’s text and image encoders to transform user queries into embeddings and perform similarity search for accurate retrieval.

Key Features:#

  • Text-to-Image Retrieval:

    Converts text prompts into embeddings and returns the most relevant images. Supports optional filters to refine search results.

  • Image-to-Image Retrieval:

    Uses a query image to find visually similar images.

  • Vector Search with Metadata:

    Performs top-k similarity search in Milvus and returns linked metadata for each result.

  • Scalable Retrieval:

    Supports large-scale datasets with fast nearest-neighbor search.

  • Integration with Milvus:

    Utilizes the Milvus vector database for efficient storage and retrieval of embeddings. Ensures high performance and scalability for large datasets.

Programming Language: Python

How It Works#

The Vector Retriever microservice provides efficient semantic retrieval over visual datasets by searching embedding vectors stored in Milvus.

  • Query Encoding:

    User input (text or image) is encoded into a vector embedding with CLIP.

  • Similarity Search:

    The query embedding is matched against indexed embeddings in Milvus to find the nearest vectors.

  • Result Generation:

    The retrieved results include metadata, similarity scores, and unique identifiers. Results are returned in JSON format for easy integration with downstream applications.

  • Result Ranking:

    Retrieved candidates are ranked by similarity score, and top-k results are returned.

  • Metadata Resolution:

    The service returns associated metadata (for example file path, source reference, or original image linkage) to provide context for each match.

Workflow#

  1. The embedding model generates text embeddings for input descriptions (e.g., “traffic jam”).

  2. The search engine searches the vector database for the top-k most similar matches.

  3. Generate results with the matched vector ids and metadata.

Learn More#