I've been benchmarking a few embedding pipelines for my Relevance AI workflows, specifically using their vector search and chat features. The goal was to balance cost, latency, and retrieval quality for a product-analytics knowledge base.
After testing, here's my current setup:
**Pipeline 1: For high-accuracy Q&A**
* **Embedder:** `text-embedding-3-large` (OpenAI)
* **Chunker:** `sliding_window` with a 512 token size and 20% overlap.
* **Results:** Top-tier for nuanced questions about user behavior metrics. The larger vector size (3072) gives great separation in the vector space. Costs more, but worth it for core features.
**Pipeline 2: For faster, cost-sensitive searches**
* **Embedder:** `embed-english-v3.0` (Cohere)
* **Chunker:** `recursive` by tokens.
* **Results:** Really impressive speed and much lower cost. I use this for broader semantic searches across support docs. Sometimes I set the `input_type` to `"search_document"` for the index and `"search_query"` at query time.
**Key finding:** The chunking strategy made a bigger performance difference than I expected. For technical docs, the sliding window overlap was crucial.
Has anyone else compared these or tried mixing embedders in a multi-vector pipeline? Curious about your benchmark results.
--ash
data over opinions