Pack RAG para PDFs y Papers de Investigación

What's in this pack

If you are a researcher, analyst, or lawyer, the bottleneck is not search — it is the PDF. Papers, contracts, filings, white papers, regulator memos. Most arrive as 1990s-era PDFs with two-column layouts, scanned pages, embedded tables, footnotes that matter more than the body. Throwing them at a general-purpose chatbot fails for the same three reasons every time: parsing is wrong, retrieval is dumb, and the model never sees the right chunk.

This pack is structured as a pipeline, not a shopping list. Each tool owns one stage, and the install order is the order data flows. Different from the PhD Researcher's Literature + Code Pack, which covers literature search and code reproduction — this pack assumes you already have the PDFs and need to actually talk to the corpus.

Install in this order

Stage 1 — Parse (turn PDFs into clean markdown)

1. Zerox — vision-model OCR for any PDF. Converts pages to images and asks GPT-4o or Claude to return clean markdown. Wins on dirty scans, two-column papers, and contracts where layout matters. The bet: a frontier vision model beats a 2018 OCR stack on hard PDFs, and you pay only when you actually run it.
2. OpenDataLoader PDF — text-first parser tuned for AI ingestion. Preserves structure (sections, tables, lists) into clean JSON or markdown. Faster and cheaper than Zerox for born-digital PDFs (papers from arXiv, recent contracts). Run this first; fall back to Zerox for the 10% that fail.
3. Surya — open-source OCR for 90+ languages. Mandatory if your corpus has Chinese, Japanese, Arabic, or Cyrillic papers. Runs locally — confidential drafts never leave your machine.

Stage 2 — Index (embed and store the parsed text)

4. Cherry Studio Knowledge Base — local RAG with native support for 50+ formats. The fastest way to drop a folder of PDFs and get a chat UI on top, all on your laptop. Start here unless you need multi-user or cloud.
5. Pinecone Assistant — managed RAG service with auto-indexing. When the corpus crosses ~10k documents or your team needs shared access, Pinecone Assistant handles ingestion, embedding, retrieval, and citations without you wiring it. Trade privacy for scale.
6. PageIndex — document index for reasoning-based RAG. Instead of flat chunk embeddings, PageIndex builds a hierarchical table-of-contents-aware index. The retrieval quality on long papers (40+ pages) is visibly better because the model can reason about where in the document an answer lives.

Stage 3 — Chat (the user-facing layer)

7. RAGFlow — deep document understanding RAG engine. The best open-source option for tables, complex layouts, and citation-grounded answers. Self-hosted, runs on Docker, includes a complete chat UI with source highlighting.
8. Kotaemon — open-source RAG document chat (the ChatPDF clone people actually keep using). Lighter than RAGFlow, easier to deploy, hot-swappable LLMs, multi-PDF chat works out of the box.

Stage 4 — Rerank and Translate

9. Cohere Rerank — boost RAG accuracy with Rerank-3. Drop in front of any retriever. The single highest-leverage 10 lines of code you can add to a RAG stack — typical relevance lift is 20-40% on noisy corpora.
10. PDFMathTranslate — translate PDF papers while preserving the original layout, equations, and figures. Essential if half your reading list is in another language and you want a side-by-side compare before feeding it to the index.

How the stages fit together

PDFs in folder
   │
   ├─ OpenDataLoader (born-digital, fast)
   │
   ├─ Zerox (dirty scans, complex layouts)
   │
   └─ Surya (non-English OCR)
         │
         ▼
   clean markdown + structure
         │
         ├─ Cherry Studio KB (local, laptop scale)
         │
         ├─ Pinecone Assistant (cloud, team scale)
         │
         └─ PageIndex (long-doc, reasoning-aware)
               │
               ▼
         ┌─────────────────┐
         │ RAGFlow         │
         │ or Kotaemon     │
         │ (chat UI)       │
         └─────────────────┘
               │
               + Cohere Rerank in front of retrieval
               + PDFMathTranslate before ingest for non-EN papers

The critical insight: most failing RAG demos fail at parse, not at retrieval. If your tables come out as Table 1 with no data, no retriever fixes that. Spend Day 1 on Stage 1; the rest gets easier.

Tradeoffs you'll hit

• Local vs cloud — Cherry Studio KB and Kotaemon run on your laptop; Pinecone Assistant ships your text to a vendor. For confidential corpora (legal, medical, M&A), stay local.
• RAGFlow vs Kotaemon — RAGFlow has the better table parser and citation UI; Kotaemon is easier to deploy and customize. Pick RAGFlow if your corpus is table-heavy (financials, scientific papers); Kotaemon for prose-heavy (legal memos, white papers).
• Zerox cost — vision-model OCR is roughly $0.01-0.03 per page on GPT-4o. A 200-paper corpus at 30 pages average runs $60-180 once. For ongoing pipelines, route only failed parses to Zerox.
• Cohere Rerank API key — adds a third-party dependency. If that's a dealbreaker, you can self-host a reranker (BGE-reranker, Jina), but the integration work is real.

Common pitfalls

• Chunk size set blindly to 512 tokens — fine for general text, catastrophic for papers where a method section runs 4000 tokens. Match chunk size to the document type.
• No source highlighting in the chat UI — researchers won't trust an answer without seeing the page. RAGFlow and Kotaemon both do this well; if you build your own UI, ship citations from day one.
• Ingesting before parsing is verified — open 5 random parsed outputs by hand before pushing 200 PDFs through the embedder. Bad parsing pollutes the index irreversibly.
• Forgetting to rerank — every team adds Cohere Rerank in week 3 after complaining about retrieval quality. Add it in week 1.