Skip to content

About

Chat with your PDFs (Arabic/English, scanned or digital): grounded RAG answers with citations highlighted in yellow on the PDF, match scores, statistics extraction. Next.js + FastAPI + AWS Bedrock + Qdrant

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

3 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PDF Insight

Upload PDFs (Arabic/English, scanned or digital), chat with them, extract every statistic, and verify each answer against highlighted sources.

Architecture

Architecture: web app and API, background ingestion, grounded chat with yellow citations

Open the interactive diagram: click any component for what it does, its source file and docs.

Ingestion (per upload, background)

flowchart LR
  U(["User<br/>drag & drop PDFs"]) -- "POST /api/documents" --> API["FastAPI<br/>magic-byte + size check"]
  API -- "original PDF" --> S3[("S3 / local disk")]
  API -- "queue" --> W["Ingestion worker<br/>thread pool"]
  W --> P["PyMuPDF<br/>words + boxes · tables"]
  W --> O["Tesseract OCR<br/>scanned pages · ara+eng"]
  W -- "large figures" --> V["Claude Sonnet 4.6<br/>chart → data points"]
  P --> N["Arabic repair<br/>ligatures · RTL cells · digits"]
  O --> N
  N --> C["Chunks ~900 chars<br/>+ anchors · cross-refs"]
  V --> C
  N --> ST["Statistics<br/>numbers · % · KPIs<br/>labelled by Claude"]
  V --> ST
  C -- "passages" --> E["Cohere Embed v4"]
  E -- "vectors" --> Q[("Qdrant<br/>filter: document_id")]
  C -- "chunks + word boxes" --> DB[("PostgreSQL")]
  ST --> DB
Loading

Chat (per question)

sequenceDiagram
  autonumber
  actor U as User
  participant UI as Next.js + PDF.js
  participant API as FastAPI
  participant BR as AWS Bedrock
  participant Q as Qdrant
  participant DB as PostgreSQL

  U->>UI: Ask a question
  UI->>API: POST /api/documents/{id}/chat
  API->>DB: Last 6 messages (follow-up context)
  API->>BR: Cohere Embed v4 (question)
  API->>Q: Top 24 passages, this document only
  API->>BR: Cohere Rerank 3.5 → best 8 sources
  API->>BR: Claude Sonnet 4.6 (tool use)
  BR-->>API: Statements + verbatim quotes per source
  Note over API: Align each quote to PDF word boxes<br/>score = 0.7 × rerank + 0.3 × fidelity<br/>≤ 45% if a number is missing from the source
  API->>BR: Cohere Rerank 3.5 (statement vs cited passage)
  API->>DB: Save question + answer
  API-->>UI: Statements, citations, boxes, %, cross-refs
  UI-->>U: Yellow highlights on the PDF · High / Medium / Low
Loading

Features

  • Drag & drop multi-file upload, 100+ page documents, OCR for scanned pages (Tesseract ara+eng)
  • Tables, figures and charts (Claude vision reads charts into text + statistics)
  • Statistics table with search/filter/sort and Excel/CSV export
  • RAG chat: answers only from the document; every statement cites verbatim quotes
  • Citations highlighted in yellow on the PDF with a match % and High / Medium / Low level
  • Resizable split view, click-to-source, clickable cross-references ("Table 3.3", "جدول ٢", "المادة ٥")
  • Login, per-user documents and chat history, private storage

Stack

Layer Tech
Web Next.js 16, React 19, Tailwind 4, PDF.js
API FastAPI, SQLAlchemy (SQLite / PostgreSQL), PyMuPDF, Tesseract
AI (AWS Bedrock) Claude Sonnet 4.6 (answers, stats labels, charts), Cohere Embed v4, Cohere Rerank 3.5
Vectors Qdrant (embedded locally, server in production)
Files Local disk or S3 (SSE / KMS)

Run locally

# API
cd backend
python3.13 -m venv .venv && .venv/bin/pip install -r requirements.txt
cp .env.example .env          # set BEDROCK_AWS_PROFILE to a profile with Bedrock access
brew install tesseract tesseract-lang
.venv/bin/uvicorn app.main:app --port 8000

# Web (new terminal)
cd frontend && npm install && npm run dev   # http://localhost:3000

API docs: http://localhost:8000/api/docs

Run with Docker

cp backend/.env.example backend/.env   # set BEDROCK_AWS_PROFILE= (empty) to use the IAM role
echo "POSTGRES_PASSWORD=$(openssl rand -hex 16)" > .env
docker compose up -d --build           # http://localhost:3000

Deploy on AWS

  • Compute: ECS Fargate or App Runner (two images), or one EC2 instance with docker compose
  • Data: RDS PostgreSQL (DATABASE_URL), S3 bucket (STORAGE_BACKEND=s3, S3_BUCKET, optional S3_KMS_KEY_ID), Qdrant container or Qdrant Cloud (QDRANT_URL, QDRANT_API_KEY)
  • IAM role: bedrock:InvokeModel (Sonnet, Embed v4), bedrock:Rerank in RERANK_REGION, S3 read/write on the bucket
  • Production settings: JWT_SECRET, COOKIE_SECURE=true, ALLOW_REGISTRATION=false once users are created, HTTPS in front (ALB/CloudFront)
  • Route /api/* straight to the API at the load balancer for very large uploads

Configuration

Variable Default Purpose
BEDROCK_AWS_PROFILE – AWS profile; empty = default credential chain / IAM role
BEDROCK_REGION / RERANK_REGION us-east-1 / us-west-2 Bedrock regions
CHAT_MODEL / STATS_MODEL / EMBED_MODEL Sonnet 4.6 / Sonnet 4.6 / Cohere Embed v4 Model IDs
DATABASE_URL SQLite in DATA_DIR PostgreSQL in production
QDRANT_URL embedded Qdrant server
STORAGE_BACKEND local local or s3
MAX_UPLOAD_MB 200 Per file
OCR_LANGUAGES / OCR_DPI ara+eng / 300 Tesseract
STATS_LLM / CHART_VISION true / true LLM labelling of statistics, chart reading

How the matching score works

Each answer statement cites sources with a verbatim quote. For every citation:

score = 0.7 × Rerank(statement, cited passage) + 0.3 × quote fidelity

  • Rerank – Cohere Rerank 3.5 relevance of the statement to the cited passage (works across Arabic/English)
  • Quote fidelity – fuzzy alignment of the quote against the PDF text (catches paraphrased or invented quotes)
  • If the statement contains a number that does not appear in the source, the score is capped at 45%
  • High ≥ 75%, Medium ≥ 50%, otherwise Low

How highlighting works

Each word's box is stored at ingestion as page coordinates (0–1). The quote is aligned to the page words, the matching words are merged into one box per line, and the viewer draws those boxes as yellow overlays on the PDF.js canvas, so they stay exact at any zoom. Arabic text is repaired (broken ligatures, RTL table cells) before matching.

Security

httpOnly session cookie (JWT), scrypt passwords, ownership checks on every document route, PDF magic-byte and size validation, CSV/Excel formula-injection guard, S3 server-side encryption.

About

Chat with your PDFs (Arabic/English, scanned or digital): grounded RAG answers with citations highlighted in yellow on the PDF, match scores, statistics extraction. Next.js + FastAPI + AWS Bedrock + Qdrant

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages