How pavankunchala/llm-learn-pk works
Project type and purpose
• This is a monorepo “LLM-Learn-PK”—a learning playground of independent AI/RAG/LLM experiments and mini-apps, including:
• Full-stack demos (e.g., resume-tailor, polygon-zone-app)
• Chainlit-based documentation agents (crewai, llama_index_doc_bot)
• Local-model demos (parking_analysis via Ollama, vision-parsing OCR)
• Data analysis scripts (csv_analyzer, customer_support)
• DSP/RAG/Quantization experiments
• Purpose: explore various LLM/RAG techniques, end-to-end pipelines, frontend/back-end integrations, and model/tooling experiments in one repo.
File structure
• Top-level folders each host a self-contained subproject.
• “resume-tailor”: React/Vite frontend + Flask backend + RAG (logs + vectorstore placeholder “rag”)
• “polygon-zone-app”: TypeScript/Vite frontend + Python WebSocket backend for video polygon drawing
• “Chainlit-apps”: Chainlit servers + vectorstore directories + crawler scripts
• “ollama/parking_analysis” & “vision-parsing”: Python test scripts + media assets + local LLM calls
• “DSP”, “Quantization”, “AIAgent-CrewAi”, “Agno-AI-Agents”: plain Python scripts and configs for experiments
• Configuration & build files present per subproject (requirements.txt, package.json, pyproject.toml, bolt config, Tailwind/Vite configs).
README for additional insights
• Highlights five featured demos with GIFs and links.
• No single unified architecture—each demo has its own stack and data flow.
• Underlying theme: LLM invocation (OpenAI, Ollama), RAG (Chroma vectorstore), retrieval pipelines, custom tools, web UIs or CLI scripts.
a. Main components (group by subproject category)
• Frontend UIs (React/Vite apps, Chainlit web UIs)
• Backend services (Flask apps, Python WebSocket servers, Chainlit agents)
• LLM services (OpenAI API, local Ollama endpoint)
• Vectorstores (Chroma directories under vectorstore/)
• Data sources (PDFs, CSVs, video streams)
• Custom tools & scripts (crew.py, crawler.py, DSP/Quantization modules)
b. Relationships & interactions
• Frontend → Backend over HTTP/WebSocket → triggers
• Backend → LLM Service calls → receives text/JSON
• Retrieval path: Backend → Vectorstore read/write → RAG loop → Backend
• Data ingest: Crawler/loader scripts → Local storage (assets/, data/) → Vectorstore
• Offline experiments: scripts reading CSV/PDF → produce reports
c. Architectural patterns & principles
• Separation of concerns: UI vs. API vs. data/model layer
• Plugin-style extensibility: custom tools directory under crew, custom_data_loader, quantization hooks
• RAG pipeline pattern: ingestion → embedding → storage → retrieve → LLM compose
• Event-driven Video processing: WebSocket feed → polygon drawing → result panel
d. Key technologies and frameworks
• Frontend: React, Vite, TypeScript, Tailwind
• Backend: Flask, Python WebSocket, Chainlit
• LLM: OpenAI API, Ollama local model
• Vectorstore: Chroma, pickle/JSON artifacts
• Utilities: PyProject/Tox for Python builds, bolt prompts, DSPy experiments
Essential architectural elements
• RAG pipeline’s stages explicitly (ingest → embed → store → retrieve → LLM).
• Integration points where experiments diverge (e.g., custom_tool vs. DSP modules).
• Every component’s responsibility is clear: who owns data, who calls whom, and where state is persisted.