Challenges We Ran Into:

1. OCR Accuracy:
   Extracting text from scanned answer sheets was challenging due to handwriting variations, low-quality scans, and non-standard formatting. Fine-tuning Tesseract OCR settings and preprocessing images was necessary to improve accuracy.

2. Embedding Precision:
   Ensuring high-quality vector embeddings for diverse and complex exam content required testing and optimizing the LLMWare embedding API. Adjusting parameters for semantic similarity searches was crucial for reliable results.

3. Database Integration:
   Efficiently storing and querying large volumes of vector embeddings in Pinecone required understanding the indexing structure and optimizing query performance for real-time evaluations.

4. Feedback Generation:
   Fine-tuning the Ollama model to generate meaningful, concise, and constructive feedback that aligned with the structured answer key was a significant challenge.

5. Multi-Language Support:
   Handling answer sheets in multiple languages (if applicable) posed challenges in both OCR and embedding generation.

6. Frontend-Backend Communication:
   Establishing seamless communication between the web interface and the FastAPI backend, especially for uploading large scanned files, required debugging API endpoints and handling file streams.

7. Time Constraints:
   Building and testing an end-to-end AI-driven system with multiple components within a limited timeframe demanded efficient task prioritization and collaboration.