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

# Embeddings and reranking

> BGE, E5, GTE, Nomic. Encoder and cross-encoder reranker fused on one GPU in a single round-trip.

An embeddings pipeline takes a list of texts and returns vector representations, optionally followed by a cross-encoder reranking pass over candidate documents, all in one HTTP round-trip. RunInfra ships the recipe with BGE encoders for embeddings and BGE or Cohere-style cross-encoders for reranking, fused on a single GPU.

## Architecture

```text theme={"dark"}
POST /v1/embeddings { input: [texts...] }
  -> BGE encoder (FP16 or FP8, batched)
  -> 1024-d vectors per text

POST /v1/{pipelineId}/rerank { query, texts }    # optional pipeline-scoped second hop
  -> BGE cross-encoder reranker
  -> Sorted documents with relevance scores
```

Both models live on the same GPU and share a CUDA stream. If you fire encoder + reranker in the same request (via a custom pipeline route), they execute back-to-back without an HTTP hop.

## What you get out of the box

* **OpenAI-compatible `/v1/embeddings`** with batched input, billing per input token
* **Pipeline-scoped `/v1/{pipelineId}/rerank`** endpoint with a text array and scored output
* **Pooled inference** sharing one GPU across both models when traffic is bursty
* **Tens of thousands of embeddings per second** on L40S with FP8 batching

## Example prompt

In [the dashboard](https://runinfra.ai/~):

```text theme={"dark"}
Build me an embeddings pipeline for English documents.
Use BGE-large-en-v1.5 plus the BGE reranker. Optimize for throughput.
```

## Quick example

```python theme={"dark"}
from openai import OpenAI

client = OpenAI(base_url="https://api.runinfra.ai/v1", api_key="YOUR_RUNINFRA_API_KEY")

resp = client.embeddings.create(
    model="your-pipeline-id",
    input=["RunInfra is a chat-native AI infrastructure platform.", "BGE is an embedding model."],
)
print(resp.data[0].embedding[:5])
```

## Models in the catalog

* **BGE** (BAAI): bge-large-en-v1.5, bge-m3 (multilingual), bge-reranker-large
* **E5** (Microsoft): e5-large-v2, e5-mistral-7b-instruct
* **GTE** (Alibaba): gte-large, gte-Qwen2-7B-instruct
* **Nomic**: nomic-embed-text-v1.5

## Deeper details

See the [models catalog](/features/models) for the full embedding model list, dimensions, and license summaries, and [runinfra.ai/use-cases/embeddings](https://runinfra.ai/use-cases/embeddings) for benchmark numbers.