I’m building a search experience over PDFs and want the best embeddings approach

For PDFs, the “best” embeddings approach is usually not just about the embedding model — it’s about how you split, clean, and retrieve PDF text. A strong default stack looks like this:

Recommended approach

1) Extract text carefully

PDFs are messy. Before embedding:

• remove headers/footers if possible
• fix hyphenation across line breaks
• preserve tables separately if they matter
• keep page and section metadata
• OCR scanned PDFs first if needed

2) Chunk by structure, not just length

Use chunks based on:

• headings / sections / paragraphs
• semantic boundaries
• token budget around 200–500 tokens per chunk, with overlap if needed

For PDFs, structure-aware chunking usually beats fixed-size chunking.

3) Use a strong modern embedding model

Good choices depend on your stack:

Open-source / self-hosted

• bge-large-en-v1.5 — strong general-purpose retrieval
• e5-large-v2 — excellent retrieval quality
• bge-m3 — good if you need multilingual + hybrid use cases
• jina-embeddings-v3 — strong and flexible

Managed APIs

• OpenAI embeddings are a strong default for quality and simplicity
• Cohere embeddings are also very good for retrieval

If you want the highest quality for search, I’d generally prefer:

• bge-m3 if multilingual / hybrid retrieval matters
• e5-large-v2 or bge-large-en-v1.5 for strong English retrieval
• a top-tier API embedding if you want easiest ops and strong out-of-the-box performance

4) Use hybrid retrieval

For PDFs, pure vector search often misses exact terms like:

• product names
• acronyms
• legal clauses
• figure/table references

Best practice:

• combine BM25 / keyword search + vector search
• merge results, then rerank

5) Add a reranker

A reranker often gives the biggest quality boost after chunking. Use a cross-encoder reranker to reorder top results from hybrid retrieval.

6) Query-time enhancements

Helpful techniques:

• query expansion / rewriting
• metadata filters by document, page, date, section
• reranking top 20–100 candidates
• answer citations to page numbers

---

Best practical default setup

If you want a solid production baseline:

• PDF parsing: structured extraction + OCR if needed
• Chunking: 300–500 tokens, section-aware, 10–20% overlap
• Embeddings: bge-large-en-v1.5 or e5-large-v2
• Retrieval: hybrid BM25 + vector
• Reranking: cross-encoder reranker
• Metadata: store page number, section, document title, table/figure flags

---

If your PDFs are tricky

Scanned PDFs

• OCR first
• use layout-aware OCR if tables/forms matter

Legal / technical PDFs

• hybrid search is especially important
• chunk by clause/heading
• preserve exact text for citations

Academic papers

• keep abstract, sections, references separate
• use section-aware chunking

Tables-heavy PDFs

• extract tables into structured text or CSV
• embed table summaries separately

---

My short recommendation

If you want the best overall results for PDF search:

1. Use structure-aware chunking
2. Use a strong embedding model like bge-large-en-v1.5 / e5-large-v2 / bge-m3
3. Combine vector + BM25
4. Add reranking
5. Preserve page and section metadata for citations

If you want, I can suggest:

• the best embedding model for your exact use case
• a reference architecture
• or a Python implementation for PDF ingestion + search.