What should I use for embeddings for PDF search and document retrieval?

For PDF search / document retrieval, the best embedding model depends on whether you want:

1. Simple semantic search
2. High-quality retrieval across lots of documents
3. Low cost / fast inference
4. On-device / open-source deployment

Good default choices

If you want the easiest strong baseline

• OpenAI text-embedding-3-small
  • Great quality
  • Cheap
  • Fast
  • Usually a very solid default for PDF search

If you want higher quality and can pay more

• OpenAI text-embedding-3-large
  • Better retrieval quality
  • Useful when search precision matters more
  • Good for complex, heterogeneous documents

Strong open-source options

• BAAI/bge-large-en-v1.5 for English
• BAAI/bge-m3 for multilingual + hybrid retrieval use cases
• intfloat/e5-large-v2 as another excellent general-purpose choice

---

What I’d recommend in practice

For most applications

Use:

• text-embedding-3-small if you want a low-cost, strong baseline
• text-embedding-3-large if quality is more important than cost

If your PDFs are mostly technical, legal, or long-form docs

Consider:

• bge-m3 or e5-large-v2
• These work well in retrieval pipelines, especially if you chunk documents well

---

Important: embeddings alone aren’t enough

For PDF search, results depend heavily on:

1. Text extraction quality

• Use a good PDF parser
• OCR scanned PDFs if needed
• Preserve headings, tables, and section structure where possible

2. Chunking strategy

• Split documents into meaningful chunks
• Common chunk size: 200–500 tokens
• Add overlap: ~10–20%
• Keep metadata like:
  • page number
  • section title
  • document name

3. Retrieval method

Best results often come from:

• Dense vector search with embeddings
• plus keyword/BM25 search
• optionally reranking with a cross-encoder or LLM reranker

This is often called hybrid search, and it performs better than embeddings alone for PDFs.

---

Quick recommendations by scenario

Small project / prototype

• text-embedding-3-small
• Vector DB: FAISS, Pinecone, Weaviate, pgvector

Best quality with OpenAI

• text-embedding-3-large
• Add reranking if possible

Fully open-source

• bge-m3
• Or e5-large-v2 for English
• Use hybrid search if you can

Multilingual PDFs

• bge-m3 is a strong choice

---

Rule of thumb

If you’re unsure, start with:

text-embedding-3-small + good chunking + metadata + reranking

Then upgrade to text-embedding-3-large if you need better retrieval accuracy.

If you want, I can also recommend:

• the best embedding model for your budget
• a PDF chunking strategy
• or a full RAG stack for PDF search.