Embedding

Turn text into vectors for search, RAG, and long-term memory.

What it is

Embeddings convert text into numerical vectors (typically []float32) that preserve semantic similarity. Once you have vectors, you can do things like similarity search, clustering, deduplication, and retrieval.

In Phero, embeddings are a building block you combine with a vector store and the RAG engine. Provider implementations live in subpackages (for example, embedding/openai).

Basic usage

You typically embed either user queries or document chunks. In the RAG examples, text is split into chunks and then embedded before it’s stored in a vector store. 

// From examples/rag-chatbot (edited for brevity)

splitter := textsplitter.NewRecursiveCharacterTextSplitter(chunkSize, chunkOverlap)
chunks := splitter.SplitText(string(fileBytes))

// Embed a chunk (or a query) to get vectors
vecs, err := embedder.Embed(ctx, []string{chunks[0]})
if err != nil {
    // handle error
}

_ = vecs // []embedding.Vector (typically []float32)

Example: infer vector size (RAG + vector stores)

Many vector stores require a fixed vector dimension for a collection/index. The RAG Chatbot and Long-Term Memory examples infer the dimension from the embedder by embedding a probe text. 

// From examples/rag-chatbot and examples/long-term-memory (edited for brevity)

// Qdrant collections require a fixed vector size. Infer it from the embedder.
vecs, err := embedder.Embed(ctx, []string{"vector size probe"})
if err != nil {
    // handle error
}
if len(vecs) != 1 || len(vecs[0]) == 0 {
    // handle error
}

vectorSize := uint64(len(vecs[0]))

Using embeddings in Phero

Embeddings are usually combined with a vector store to power RAG:

See the examples: RAG Chatbot and Long-Term Memory.

Run an example

These examples exercise embeddings as part of a full retrieval pipeline. Follow each example’s README for any required services and provider setup. 

# from repo root
go run ./examples/rag-chatbot -file ./README.md

go run ./examples/long-term-memory

Related packages