Myanmar OCR Pipeline

📄 Upload Document

Select a Myanmar document image (PNG, JPG, TIFF) or PDF. Supported formats: PNG, JPG, JPEG, TIFF, TIF, PDF. Max file size: 50MB.

📤

Drop file here or browse

PDF · PNG · JPG · TIFF

—

— —

After uploading, click "Start Processing" to run the full integrated pipeline: OCR (7 agents) → Metadata → Dataset → RAG (chunking + embedding). All progress is displayed in real-time. Total pipeline takes ~2-5 minutes.

🔄 Full Pipeline Progress ? —

Waiting for pipeline to start...

⏳ OCR Pipeline

→

⏳ Metadata

→

⏳ Dataset

→

⏳ RAG / Chunks

⚡ OpenClaw Pipeline Monitor ? —

7 agents run sequentially. Each dot shows current status. Duration per agent displayed in the table below.

	Agent	Duration	Status / Error

🔬 Detailed Pipeline Trace ?

Expand each step to see the actual text output at that stage of processing. Trace how text transforms from raw OCR through to cleaned output.

🏷️ Metadata Trace ?

Each metadata agent's input and output. Expand to see the extracted values and validation results.

🏷️ Document Metadata ?

Metadata is generated using a local LLM (Qwen2.5:7b via Ollama) after OCR completes. Click "Generate Metadata" to analyze the document and extract structured information.

📊 Dataset Builder ?

Transform processed documents and metadata into structured datasets for fine-tuning LLMs, synthetic data generation, and knowledge management. Datasets are stored with versioning in PostgreSQL.

📈 Dataset Dashboard ? ↻ Refresh

Real-time statistics for all datasets generated by the Dataset Builder pipeline.

—

Total Datasets

—

Total Records

—

Success Rate

—

Failure Rate

📚 Knowledge

0

💡 Instruction

0

📝 Summary

0

🏷️ Classification

0

—

Avg Validation Score

—

Failed Validations

—

Duplicate Detections

🧠 RAG Pipeline — pgvector + Chunking + Semantic Search ? ↻ Refresh

Phase 4 — RAG Pipeline: After OCR & Metadata complete, click "Ingest to RAG" to chunk the cleaned text, generate embeddings (Ollama all-minilm, 384d), and store in pgvector. Then use the Chat interface for semantic Q&A.

—

Total Chunks

—

Documents Indexed

—

Indexed / Completed

—

Embedding Model

—

🔍 Semantic Chunk Search

Enter a query to search across all indexed chunks using cosine similarity.

📋 RAG Workflow:


Clean Text → [Chunking: 1000 chars, 200 overlap] → N chunks →
[Ollama all-minilm embedding (384d)] → vectors →
[pgvector INSERT + ivfflat index (cosine, 100 lists)] →
Semantic search (cosine similarity) → [Ollama Qwen2.5:7b LLM] → Answer

⬇️ Export JSONL for Fine-Tuning ? ↻ Refresh

Select datasets to export as JSONL. Combine all selected datasets into a single file for training, or download individually. Compatible with ChatML format for fine-tuning.

—

Loading datasets...

⚠️ Danger Zone

Permanently delete all documents, metadata, datasets, RAG chunks, chat history, and uploaded files. This action cannot be undone.

🔍 Search Documents ?

Search through previously processed documents. Use filters to narrow results by topic, document type, or region. Results show metadata including title, type, and confidence score.

Enter a query and click Search.

📋 History Log ? ↻ Refresh

Shows recent documents. Status indicators: completed ✅, processing ⏳, failed ❌. Click any row to view full OCR results and metadata.

No documents uploaded yet.

Myanmar OCR Pipeline - MVP 8

Dataset: