Recommendation (short): Build a 3-stage chain: 1) chunk + extract structured facts per chunk, 2) merge + dedupe evidence across chunks, 3) synthesize a short brief with inline attributions. Use ChatGPT (or Claude for extra-long context/careful analysis) to run the extraction and synthesis prompts. Below are chunking patterns, ready-to-use prompt templates, decision criteria, a practical checklist, and best-for / avoid-if guidance.

Chunking patterns (pick one):
- Page-based: one chunk = one PDF page. Best when pages are dense and page numbers matter.
- Section-based: split by headings (Introd, Methods, Results). Best when PDFs have clear structure.
- Fixed-token windows: 800–1,200 tokens per chunk with 150–250 token overlap. Best when layout is inconsistent or OCR text is noisy.
Naming convention: assign each chunk a unique ID like DOC03_PG12 or DOC07_SEC2 and include source filename and page range in metadata.

Extraction prompt (per chunk) — system + user combined template:
"You are an assistant that extracts structured facts from a text chunk. Return JSON only. Input fields: chunk_id, source_filename, page_range, text.
Return: {"chunk_id":..., "facts":[{"fact_id":"unique","claim":"short sentence","evidence":"quoted excerpt","page":"p#","confidence":0-1,"tags":[topic keywords]}], "issues": ["OCR errors or unclear passages"], "summary":"2-3 sentence summary"}
Extract all claims, data points, direct quotes ( [source chunk ids + quoted evidence], 3) rate overall confidence and note contradictions.
Output JSON: {"claims":[{"claim":"...","sources":["DOC01_PG3"],"evidence":["quote..."],"confidence":"high/med/low","contradictions":[]}], "contradictions_summary":"..."}
Prioritize explicit quotes and numeric data; do not invent page numbers.

Final brief prompt (synthesize):
"Produce a concise brief (400–800 words) with: 1) Executive summary (2-3 bullets), 2) Top findings (numbered, each with 1-2 sentence synthesis + inline attributions like [DOC03_PG12]), 3) Important quotes (with citations), 4) Evidence table (claim -> sources), 5) Methodology note (how chunks were made). Keep language neutral and include a bibliography of documents (filename -> citation)."

Decision criteria (choose tool/parameters):
- Use Claude if you need very long single-context merges or higher safety/nuance; use ChatGPT for speed, cost-effectiveness, and broad availability.
- Use smaller chunk sizes if OCR quality is low; larger if structure is clear.
- Team size & skill: manual spot-checking required by 1 reviewer per 10 docs.
- Budget: more paid tokens/models reduces need for aggressive chunking.

Practical checklist (step-by-step):
1) OCR PDFs with page markers. 2) Extract text and split into chunks (choose pattern). 3) Assign chunk IDs and metadata. 4) Run extraction prompt for each chunk (parallelize). 5) Run merging prompt on all extraction outputs. 6) Resolve contradictions manually for critical points. 7) Run final brief prompt. 8) Spot-check 10% of claims against source images. 9) Add bibliography and provenance table. 10) Save both full evidence matrix and concise brief.

Best-for: teams needing traceable, auditable briefs from many documents. Avoid-if: you need instant, inference-free legal opinions without human review.

Tools mentioned: chatgpt, claude

Compare ChatGPT and Gemini