Which embedding model should I use for classification on my internal docs?

For classification on internal documents, the best embedding model is usually:

Short answer

Use a strong general-purpose text embedding model from a reputable provider, then train a simple classifier on top.

Good defaults:

• OpenAI: text-embedding-3-large for best quality, or text-embedding-3-small for lower cost
• Cohere: embed-english-v3.0 / embed-multilingual-v3.0
• Sentence Transformers (open source): all-mpnet-base-v2 is a solid baseline, but newer retrieval-oriented models may be better depending on your data

What to pick

If you want best performance

• text-embedding-3-large
• Best when accuracy matters more than cost
• Works well on messy, short, and medium-length enterprise text

If you want a cheaper option

• text-embedding-3-small
• Often good enough for internal classification tasks
• Much cheaper and faster

If you need on-prem / no external API

• Use a strong open-source embedding model, e.g.:
  • bge-large-en-v1.5
  • e5-large-v2
  • all-mpnet-base-v2 as a simple baseline

Important note for classification

Embeddings are usually not the classifier themselves. The usual recipe is:

1. Embed each document
2. Train a classifier on those vectors
   • logistic regression
   • linear SVM
   • XGBoost
   • small MLP
3. Evaluate on a held-out set

This often outperforms prompting a language model directly for stable document classification.

Practical recommendation

If you have no constraints:

• Start with text-embedding-3-large
• Train a logistic regression or linear SVM
• Compare against text-embedding-3-small for cost/perf

If your docs are long

For long internal docs, consider:

• embedding the title + summary + key sections
• chunking the document and pooling chunk embeddings
• using only the most relevant section for classification if labels depend on specific content

If you tell me:

• document type,
• number of labels,
• language(s),
• whether you need on-prem,
• and approximate dataset size,

I can recommend a more specific model and training setup.