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.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
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.
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-endpoints when reranking requires a dedicated deployment
Confirm that the user needs vectors or retrieval, not direct generation.
Choose the embedding model and batch shape.
Generate embeddings for corpus and query paths consistently.
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.
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.
