Tensor TownISSUE #98 of 120

embeddings models · openai · cohere · bge

NeuraVSThe Overfit Ogre

Neura saysEmbedding models (OpenAI, Cohere, BGE) turn text into vectors for search and RAG.

Dedicated embedding models are optimized to map text to vectors where semantic similarity = closeness — the retrieval engine behind search, RAG, clustering, and recommendations. Options range from API services (OpenAI, Cohere) to strong open models (BGE, E5). Key choices: dimension (storage/speed), domain fit, and matching the same model for indexing and querying. Good embeddings are half of good retrieval.

Power-ups you unlock

Text → similarity-preserving vectors
Engine behind search, RAG, clustering
API (OpenAI/Cohere) or open (BGE/E5)
Use the same model to index and query

The Overfit Ogre attacks — common mistakes

Mixing embedding models for index vs query
Ignoring domain fit of the model
Over-paying for dimensions you don’t need

Boss battleExplain why index and query must use the same embedding model.

Example code

<!doctype html><html><head><meta charset="utf-8"></head>
<body style="background:#06040d;color:#e6e0ff;font-family:monospace;padding:20px"><pre>docs → embeddings (model X) → vector DB
query → embedding (model X) → nearest docs
mismatched models → broken similarity</pre></body></html>

▶ Open the interactive comic issue

‹ Video Generation · Sora · Veo Vector Dbs · Pinecone · Weaviate · Chroma ›