CLAUDE.md Template for High-Fidelity PDF Chat & Document RAG

What is this CLAUDE.md template for?

This CLAUDE.md template directs your AI coding assistant to design PDF chat applications with an intense focus on layout awareness and semantic correctness. Most generic AI routines process PDFs as continuous, unformatted strings of text, which destroys structural meaning like table headers, nested sections, footers, and page numbers.

This template locks down strict rules for multi-modal layout parsing, converting embedded charts and tables into clear Markdown formats, maintaining rigorous document coordinate mapping, and enforcing absolute grounding requirements with verifiable page-specific citations.

When to use this template

Use this template when building document analytics tools, financial auditing platforms, legal document search systems, invoice extraction engines, or any chat-with-PDF system where missing an embedded table or hallucinatory reference could compromise application reliability.

Recommended engineering ingestion flow

[Raw PDF Upload] 
       │
       ▼
[Layout-Aware Parser] ──► (Isolate headers, footers, sidebars)
       │
       ▼
[Table/Chart Extractor] ──► (Serialize structures into clean Markdown tables)
       │
       ▼
[Metadata Coordinate Tag] ──► (Stamp nodes with exact page and line numbers)
       │
       ▼
[Vector Store / Grounded RAG] ──► (Execute retrieval utilizing strict source grounding)

CLAUDE.md Template

# CLAUDE.md: High-Fidelity PDF Processing & Chat Engineering Guide

You are operating as an Expert Document Extraction & Semantic Search Engineer specializing in advanced PDF structural parsing, multi-modal ingestion, and strictly grounded conversational RAG layers.

Your mandate is to build zero-hallucination document query workflows that maintain pristine layout awareness and explicit page source continuity.

## Core Extraction Principles

- **Layout-Aware Parsing**: Never process document streams as unstructured plain text strings. Utilize layout-aware extraction models (e.g., LlamaParse, Marker, PyMuPDF with structured flags) to isolate headers, footers, sidebars, and structural columns.
- **Table & Chart Fidelity**: Embedded tables and data graphics must be extracted cleanly and serialized into explicitly formatted Markdown tables or structural JSON fields. Never let tabular data collapse into mangled inline phrases.
- **Rigid Grounding & Citations**: Responses generated by the synthesis engine must be explicitly anchored to source nodes. Enforce the output format to append verified citations including file names, page numbers, and exact matching text snippets.
- **Asynchronous Chunking Processing**: For large multi-page uploads, execute page extraction tasks concurrently using async tasks to optimize processing queues and system throughput.

## Code Construction Rules

### 1. Ingestion, OCR, & Chunking Protocols
- Store extracted content using structural node formats that explicitly track positioning vectors (`file_id`, `page_number`, `chunk_index`, `bbox_coordinates`).
- Implement semantic text splitting patterns (e.g., splitting by document header elements or markdown structural blocks) rather than blind character count offsets to avoid breaking logical sentences.
- Configure advanced document filters to purge repetitive header and footer strings from vector indexes to avoid indexing noisy redundant metrics.

### 2. Retrieval, Metadata Filtering, & Verification
- Route queries safely. Always combine user semantic query embeddings with explicit document visibility tags (`document_id`, `tenant_id`) within vector database filters.
- When tabular lookups are targeted, employ a routing layer that switches the query directly to an extracted metadata markdown matrix engine rather than relying on pure vector similarities.

### 3. Response Generation & Grounding Guards
- Construct system prompt layouts that strictly forbid the LLM from relying on internal historical training knowledge. If the context fragments do not contain clear validation answers, output an explicit "Information not found" message.
- Mandate structured token matching: verify that every citation page reference returned by the completion model corresponds directly to an existing ID index inside the retrieved context packet before serving the payload.

## Performance Optimization & Testing
- Write backend tracking loops to log extraction latencies, page counts, and failure rates from OCR providers.
- Implement mock integration testing files using simple multi-column and tabular PDF fixtures to verify document node coordinate integrity across systemic refactors.

Why this template matters

Chat-with-PDF applications are notoriously prone to hidden retrieval failures. Standard AI routines often slice chunks directly through the middle of an important financial table, scrambling the rows and causing the final model to hallucinate numbers completely. They also struggle to output honest citations, often guessing page numbers randomly when under pressure.

This configuration completely eliminates these systemic failures by mandating structural layout parsing tools, explicit markdown table tracking rules, and programmatic confirmation hooks that verify every single page citation before it ever leaves the API boundary.

Recommended additions

• Incorporate clear instruction sets for handling scanned, image-only documents using heavy cloud OCR pipelines (e.g., AWS Textract or Azure Document Intelligence).
• Add targeted guidance for rendering exact text-highlighting bounding boxes (BBox fields) directly within client-side PDF viewer components.
• Define caching protocols using Redis or local object stores to prevent costly re-parsing cycles on duplicate document hash uploads.
• Include explicit rules for multi-file cross-document synthesis workflows, specifying how conflicting entity terms should be resolved.

FAQ

How does this template handle multi-column research papers or financial reports?

It explicitly forces the AI assistant to adopt layout-aware extraction libraries (like LlamaParse or structured PyMuPDF rules) that read text sequentially down columns instead of incorrectly scanning across the entire horizontal width of a page.

Can this template be used with standard open-source parsers like PyPDF?

Yes, but it guides the assistant to configure those parsers defensively, wrapping text with structural metadata tags, sorting extraction coordinates, and isolating bounding layouts to minimize semantic distortion.

Why does this blueprint focus so heavily on table serialization?

Standard vector lookup fails heavily on loose tabular text. Serializing tables into crisp Markdown formats or structured JSON keys allows the context engine to preserve vertical and horizontal row-column alignments, enabling the LLM to read data arrays flawlessly.

How are hallucinatory page citations prevented?

The code design rules demand that the AI assistant build a programmatic cross-check layer. This layer takes the citations generated by the model and cross-references them against the actual retrieved node indexes, dropping any untrusted or unverified references immediately.

About the author

Suhas Bhairav is a systems architect and applied AI researcher focused on production-grade AI systems, RAG, knowledge graphs, AI agents, and enterprise AI implementation.