I’m unhappy with Cohere embeddings on my support corpus

Sorry to hear that. “Unhappy” with embeddings can mean a few different failure modes, and the fix depends on which one you’re seeing.

Common issues with support corpora embeddings:

1. Retrieval misses obvious docs

   • Queries use different wording than tickets/docs.
   • The embedding model isn’t handling domain-specific language well.

2. Too many near-duplicate results

   • Support articles may be repetitive, causing results to collapse around generic text.

3. Bad chunking

   • Chunks are too large, too small, or split across helpful context boundaries.

4. Poor semantic separation

   • Similar but distinct issues (billing vs login vs device activation) cluster too closely.

5. Short query problem

   • User queries are terse (“can’t sign in”), which can be hard to embed well without query expansion or reranking.

6. Language / jargon mismatch

   • Product names, error codes, and internal terminology can confuse general-purpose embeddings.

What usually helps:

• Evaluate with a small labeled set

  • Pick 50–200 real queries and mark the correct docs.
  • Measure Recall@k / MRR before changing anything.

• Use better chunking

  • Try chunk sizes around 200–500 tokens with overlap.
  • Keep titles, headings, and error codes attached to chunks.

• Add reranking

  • Use embeddings for candidate retrieval, then a reranker for final ordering.
  • This often gives a noticeable lift on support search.

• Normalize and enrich text

  • Include ticket subject + body + product + tags + FAQ title.
  • Expand acronyms and preserve error codes.

• Try a different embedding model

  • Especially one tuned for retrieval and short queries.
  • Cohere may be fine, but your corpus/query mix may fit another model better.

• Hybrid search

  • Combine lexical search (BM25) with embeddings.
  • Support content often benefits a lot from exact-match signals like error codes and product names.

If you want, I can help you diagnose this more concretely. Send me any of:

• 3–5 example queries
• the top results you’re getting
• the result you expected
• how you chunked the docs
• which Cohere embedding model you’re using

Then I can suggest specific changes.