> ## Documentation Index
> Fetch the complete documentation index at: https://aidocs.ethanbragdon.icu/llms.txt
> Use this file to discover all available pages before exploring further.

# Embeddings API Reference

> Understand Starfire AI embedding resources, model eligibility, input batching, vector metadata, metering, and the difference between embeddings and managed Knowledge.

# Embeddings API reference

Embeddings convert supported input into vector representations that applications can use for semantic similarity and retrieval.

## Resource

The Starfire v1 architecture includes an embeddings resource family. Confirm exact route, model IDs, input limits, batching behavior, and response fields in the active OpenAPI/Developer Portal.

## Embeddings vs Knowledge

Use the embeddings API when **your application manages the vector index and retrieval system**.

Use Starfire Knowledge when you want Starfire to manage ingestion, chunking, indexing, permissions, and retrieval around a persistent Knowledge Base.

## Model eligibility

Only embedding-capable models/routes can be used. Do not send an ordinary conversational model ID and assume Starfire will reinterpret it automatically.

## Input

The active contract can support one or more text inputs per request subject to size and batch limits.

Normalize your own application data consistently before embedding if you need reproducible similarity behavior.

## Output metadata

Applications should retain enough metadata to associate a vector with its source record. A vector without a stable source reference is difficult to update, delete, or explain later.

## Metering

Embedding usage can contribute to Starfire developer usage and credits independently from conversational generation.

## Security

Embedding vectors can encode information derived from private source material. Treat the resulting index as protected data appropriate to the source—not as automatically safe public metadata.

<Note>
  Starfire can change or add embedding providers/models. Applications should store the embedding model/version used for an index so incompatible vector spaces are not mixed accidentally.
</Note>