feat: #592 Abstract VLM providers into an extensible Strategy Pattern

[suhaniiz]

📋 PR Checklist

Thank you for contributing to PDF-Assistant-RAG! 🎉
Please fill out this template before submitting. PRs without it filled in will be closed.

---

🔗 Related Issue

Closes #592

---

📝 What does this PR do?

Decouples the core image captioning pipeline from individual vendor integrations by implementing a Strategy Pattern.

• Introduced a BaseVisionProvider abstract interface.
• Refactored the previous inline OpenAI placeholder logic into a standalone, production-ready OpenAIVisionProvider strategy (including base64 encoding preparation for standard multimodal payload handling).
• Introduced a central VISION_PROVIDER_REGISTRY dict factory mapping to allow seamless plug-and-play scaling for future models (e.g., Gemini, Claude, Ollama) without polluting core operational files, adhering strictly to the Open-Closed Principle.

---

🗂️ Type of Change

• 🐛 Bug fix
• ✨ New feature
• 🔧 Refactor / code cleanup
• 📝 Documentation update
• 🎨 UI / styling change
• ⚙️ CI / tooling / config change
• 🧪 Tests

---

🧪 How was this tested?

• Verified code changes sit strictly within the registry mapping and fallback handlers.
• Tested fallback logic to ensure unconfigured or failing provider runs gracefully chain downstream to the local OCR pipeline (_ocr_caption) without breaking chunk execution.

---

📸 Screenshots (if UI change)

N/A

---

⚠️ Anything to flag for reviewers?

The integration remains completely backwards-compatible; if no VLM engine string matches the registry, it safely skips ahead to local OCR execution as before.

---

✅ Self-Review Checklist

• My branch is based on dev, not main
• I have not added any secrets / API keys
• I have not modified main branch or any HuggingFace deployment config
• My code follows the existing style (no unnecessary formatting changes)
• I have updated relevant docs / comments if needed

[suhaniiz]

Hi @param20h 👋

As the original author of issue #592, I wanted to submit my official implementation for this feature as part of GSSoC '26.

While I notice another PR was opened alongside this, my implementation explicitly fulfills the architectural requirements outlined in the issue description. I have fully abstracted the VLM logic out of the core execution loop into a strictly decoupled BaseVisionProvider interface and a VISION_PROVIDER_REGISTRY factory. Additionally, the concrete OpenAIVisionProvider strategy has been updated to support modern, production-ready base64 image data URI encoding for real multimodal execution, rather than relying on standard text prompts or outdated endpoints.

I would highly appreciate it if you could review this solution, as it ensures long-term extensibility for future providers (like Gemini or Claude) without breaking the core pipeline. Thank you! 🚀