Principle:Togethercomputer Together python Result Integration
Overview
Result Integration is the pattern for combining embedding-based retrieval and reranking results into a final ranked document list for application use in the Together Python SDK workflow.
Description
Result integration is the user-defined step of taking the outputs from embedding generation and/or document reranking and producing the final document set for downstream applications. This includes mapping reranked indices back to original documents, applying relevance score thresholds, combining vector search scores with reranking scores, and feeding the final ranked list into downstream applications such as LLM context windows for RAG.
The Together Python SDK provides the raw outputs (embedding vectors from Embeddings.create() and relevance-scored indices from Rerank.create()), but the logic for interpreting, filtering, and combining these outputs is the responsibility of the user.
Usage
Use result integration after embedding generation and/or reranking to produce the final document set for your application. Common integration points include:
- RAG context assembly -- Select the top-k reranked documents and concatenate them as context for an LLM prompt
- Search result display -- Map reranked indices back to full document metadata for rendering in a search UI
- Score-based filtering -- Apply a relevance score threshold to remove low-confidence results
- Multi-signal fusion -- Combine embedding similarity scores and reranking relevance scores for a more robust final ordering
- Diversity-aware selection -- Ensure the final result set covers diverse topics rather than redundant similar documents
Theoretical Basis
In RAG pipelines, result integration bridges retrieval and generation by selecting and ordering the most relevant context. The key theoretical considerations are:
- Score thresholding -- Relevance scores from reranking models are calibrated per-query, meaning a threshold (e.g., 0.5) can meaningfully filter out irrelevant documents. This prevents low-quality context from being injected into LLM prompts, which can degrade generation quality.
- Reciprocal rank fusion (RRF) -- When combining multiple retrieval signals (e.g., BM25 keyword search + embedding similarity + reranking scores), reciprocal rank fusion provides a principled way to merge ranked lists. The RRF score for a document is the sum of
1 / (k + rank)across all retrieval systems, wherekis a constant (typically 60).
- Context window management -- LLMs have finite context windows. Result integration must select and order documents to maximize information density within the available token budget. This may involve truncating documents, selecting only the most relevant passages, or ordering documents by relevance.
- Diversity-aware selection -- Returning only the top-k most relevant documents may produce redundant results. Maximal Marginal Relevance (MMR) balances relevance with diversity by penalizing documents that are too similar to already-selected ones.
- Index mapping -- Reranking results contain indices relative to the input document list, not the original corpus. Maintaining a mapping from reranked indices to original document identifiers (e.g., database IDs) is essential for connecting retrieval results back to the full application data model.
Metadata
| Property | Value |
|---|---|
| Principle | Result Integration |
| Domain | NLP, Information_Retrieval, RAG |
| Workflow | Embeddings_And_Reranking |
| Related Concepts | Reciprocal Rank Fusion, Score Thresholding, Context Window Management, Maximal Marginal Relevance |
| Implementation | Implementation:Togethercomputer_Together_python_Result_Integration_Pattern |