agentsclimarketplace
Sign in

Together embeddings

Skill togethercomputer/skills/skills/together-embeddings

Dense vector embeddings, semantic search, RAG pipelines, and reranking via Together AI. Generate embeddings with open-source models and rerank results behind dedicated endpoints. Reach for it whenever the user needs vector representations or retrieval quality improvements rather than direct text generation.From its SKILL.md

Install
npx -y skills add togethercomputer/skills --skill together-embeddings

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

3.8 KB, as published. Nobody here has run it

Together Embeddings & Reranking

Overview

Use this skill for semantic retrieval components:

  • create embeddings
  • batch embeddings
  • build retrieval or RAG pipelines
  • rerank retrieved candidates

This skill is for retrieval plumbing, not for the final language-model response itself.

When This Skill Wins

  • Build vector search or semantic similarity features
  • Add embedding generation to a data pipeline
  • Improve retrieval quality with reranking
  • Assemble a retrieval stage before calling a chat model

Hand Off To Another Skill

  • Use together-chat-completions for the final answer-generation step
  • Use together-batch-inference for very large offline embedding backfills
  • Use together-dedicated-model-inference when reranking requires a dedicated deployment

Quick Routing

Workflow

  1. Confirm that the user needs vectors or retrieval, not direct generation.
  2. Choose the embedding model and batch shape.
  3. Generate embeddings for corpus and query paths consistently.
  4. Retrieve candidates. An in-memory cosine-similarity store works for prototyping and small corpora (see semantic_search.py). Use a dedicated vector database for production scale.
  5. Rerank only when the extra latency and endpoint requirement are justified. When no dedicated rerank endpoint is available, cosine-similarity ranking is a reasonable fallback.

High-Signal Rules

  • Python scripts require the Together v2 SDK (together>=2.0.0). If the user is on an older version, they must upgrade first: uv pip install --upgrade "together>=2.0.0".
  • Keep embeddings and reranking conceptually separate; rerank is a second-stage precision step.
  • Reranking in this repo assumes a dedicated endpoint. Do not promise serverless rerank unless the product changes. When no endpoint is available, fall back to cosine-similarity ranking.
  • The embedding model has a 514-token context limit. Chunk longer documents before embedding.
  • The rag_pipeline.py example demonstrates retrieval plus generation; treat generation as a hand-off to chat completions.
  • Preserve model consistency across indexing and querying.

Resource Map

Official Docs

Keep looking

Skills are one crate of 627,801. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.