Skip to main content

Vector Search and RAG with ApertureDB

ApertureDB can store and manage multimodal data, including text, documents, images, videos, blobs, clips, frames, and their associated embeddings. With its support for indexing embeddings of any dimensions, perfoming K-near neighbor (KNN) search and classification, ApertureDB offers all the necessary support as a vector database.

ApertureDB TermEquivalent Elsewhere
Descriptor classcollection, vector store
Vector indexsearch engine, dimensions, distance metrics
Descriptorembedding, feature vector

Descriptor classes and vector indexes​

A descriptor class groups embeddings that belong to the same vector search space. For example, an OpenFace class can hold 128-dimensional face embeddings and their metadata. Use CreateIndex with target: "descriptor", the class name, property: "_blob" and one engine as kind. For example, an index with kind: "hnsw" and params: {"dimensions": 128, "metric": "CS", "M": 32} builds an HNSW graph for cosine similarity. Other index kinds are flat, ivf, and acorn. To use multiple engines on the same class, create a separate index for each one.

Each nearest-neighbor query selects an engine and a compatible metric. See the descriptor index parameter tables for dimensions, metrics, indexing modes, and engine-specific settings.

Descriptors or Embeddings​

Unimodal or multimodal embeddings are always added to an explicit descriptor class, along with metadata properties and labels.

A blob must be provided in the array of blobs, which contains the descriptor's values. The blob is an binary array of 32-bit floating point values. The size of the blob, in bytes, must be dimensions*4.

Quickstart Notebook

Vector Search Quickstart — create a vector index, add vectors, run KNN search, and use the Python SDK Descriptors wrapper. The quick start also links to notebooks showing vector search support for various data types like images, videos, documents, audio.

Application Examples​

Some examples of ApertureDB being used as a vector database:

Building RAG Pipelines​

MMR (Maximal Marginal Relevance) reranking is implemented in the Python SDK via Descriptors.find_similar_mmr and is utilized in the LangChain and LlamaIndex integrations.

Graph capabilities let you scope retrieval to a subgraph — for example, restricting responses based on user access permissions for secure RAG or building real-world RAG applications.

See the Building RAG Pipelines guide for a full pipeline walkthrough.

ApertureDB vs. Other Vector Databases​

ApertureDB delivers 2–10x higher KNN throughput compared to Pinecone, Weaviate, Qdrant, Chroma, Lance, and Milvus across embedding sizes from 96 to 4096 dimensions, with sub-10ms query latency for typical RAG workloads. Pricing does not vary by dimension count, number of embeddings, or query volume.

See the full benchmarks page, or request the performance whitepaper (access required).

Demos​

How ApertureDB powers real-world multimodal AI applications.
Face similarity search combining metadata filters and KNN search on the CelebA dataset.

The MCP Server workflow exposes ApertureDB directly to MCP-compatible agents (Claude, Cursor, etc.) without writing retrieval code manually.

Expose ApertureDB as a tool for AI agents via the Model Context Protocol.
Crawl a website and build a RAG question-answering agent using the ApertureDB Workflows UI.
Extract text from PDFs, chunk, embed, and store as descriptors in ApertureDB — no code required.
Generate and search video frame embeddings using the ApertureDB Workflows UI.
Run face detection and similarity search using the ApertureDB Workflows UI.

For more videos, visit the ApertureData YouTube channel.

References​