DEPLOY / RAG

Answer from your own documents

Embed, retrieve, rerank and answer on your own models. Fine-tune the retriever with a retrieval file of query and passage pairs, and the pipeline improves from your own corpus.

01Indexembed the corpus once
02Retrievenearest by vector
03Rerankby relevance
04Answerfrom what was retrieved
What it costs to readmetered on tokens read
$0.03embedding, per 1M tokens
$0.09reranking, per 1M tokens
Embedding models
4
Rerankers
2
Fine-tune the retriever
a retrieval file

Four embedding models, two rerankers.

Six models from 300M to 8B parameters, all trainable with LoRA. Pick a bi-encoder for embedding and a cross-encoder for reranking.

Fine-tune the retriever.

A retrieval file of query and passage pairs trains the embedding or rerank model on your domain. A few thousand labelled pairs move recall substantially.

Metered on tokens read.

Embedding and reranking calls are priced per million tokens read, from $0.02. The models return a vector or a score, so there is no output token charge.

How it works

How retrieval-augmented generation works

Embed the corpus once

Turn every document into a vector with an embedding model. Pass an array to embed a batch in one call, up to 256 inputs. An input longer than the model's context length is rejected rather than silently truncated, so a document that will not fit is a 400 you can act on.

Embeddings API reference
Results

Results from LLMs powered by Nucleus

Same day

from site reading to client report

9

systems on one thread

6

vendors under review for Q4 renewal

Use cases

Retrieval in the field

Field engineering

Cable survey readings processed while the crew was still on site

Same dayfrom site reading to client report
9systems on one thread
Procurement

Six vendor contracts ranked before the renewal meeting

6vendors under review for Q4 renewal
3axes: cost, quality, response
IT operations

Every incident past its four-hour SLA, escalated by name

8incidents past their four-hour SLA
6.2 hrsaverage time to resolution
At a glance

What the pipeline includes

Embedding

  • 4 bi-encoder models, 300M to 8B parameters
  • Up to 4096 dimensions, Matryoshka truncation supported
  • Batch up to 256 inputs per call
  • Context from 2K to 32K tokens
  • float or base64 encoding

Reranking

  • 2 cross-encoder models, 568M and 8B parameters
  • Up to 256 documents per call
  • top_n filtering in the response
  • 100+ languages (BGE Reranker v2 M3)
  • Scores comparable within one call

Fine-tuning

  • Retrieval files of query and passage pairs
  • All 6 models trainable with LoRA rank up to 64
  • Training from $0.03/Mtok (EmbeddingGemma 300M)
  • Checkpoints exportable as adapters or merged weights
Pricing

Embedding tokens

Embedding tokens

$0.02per million tokens

EmbeddingGemma 300M inference. All models priced per million tokens read.

Pick your path

Pick your path

Harness

Permissions, scoped keys and audit

Explore Harness

On-Premise Deployment

Our GPU servers, your site

Explore On-Premise Deployment

Tool use

Your private APIs as skills

Explore Tool use

MCP

Your model in your tools

Explore MCP

Webhooks

Signed events to your systems

Explore Webhooks