AIProjects

A collection of small, independent AI/agentic projects, grouped by the framework or pattern they explore (RAG, multi-agent orchestration with CrewAI, AutoGen-style microservices + n8n). Each top-level folder is a category that can hold multiple projects; every project lives in its own subfolder with its own requirements.txt, its own README, and no shared dependencies, so it can be run entirely on its own.

Repository layout

AIProjects/
├── RAG/
│   ├── pdf_qa_assistant/         # PDF Q&A app — LangChain + FAISS + Gradio
│   └── personal_learning_rag/    # Local semantic search + daily-fact RAG over your own study materials
├── CrewAI/
│   └── agentic_rag_router/       # Agentic router/retriever RAG system (Jupyter notebook)
└── AutoGen/
    └── linkedin_automation/      # AutoGen-style FastAPI microservice + n8n workflow

To add a new project, create a new subfolder under the relevant category (RAG/, CrewAI/, AutoGen/, or a new category folder if it doesn't fit any of these) with its own code, requirements.txt, and README, then add a section for it below.

Prerequisites

• Python 3.10+
• pip
• An OpenAI API key — required by RAG/pdf_qa_assistant and CrewAI/agentic_rag_router; optional for AutoGen, which can run fully offline in mock mode; not needed at all for RAG/personal_learning_rag (fully local via Ollama)
• Tavily API key (free tier) — only for CrewAI's web search path
• Ollama + Tesseract OCR — only for RAG/personal_learning_rag, see its own README for setup

Each project keeps its own requirements.txt and its own virtual environment, since they don't share dependencies. From the repo root:

cd <project-folder>
python3 -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate
pip install -r requirements.txt

---

RAG / pdf_qa_assistant — PDF Q&A Assistant

What it is: A Gradio app that lets you upload PDF documents, indexes them (chunk → embed → FAISS vector store), and answers questions grounded only in the uploaded documents (no hallucinated answers outside the PDFs). Two sample PDFs (Company_Employee_Handbook_1.pdf, IT_Support_Guide.pdf) are included to try it out. See RAG/pdf_qa_assistant/README.md for how it works internally.

Stack: LangChain, OpenAI (gpt-4o-mini + text-embedding-3-small), FAISS, Gradio.

Run it:

cd RAG/pdf_qa_assistant
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

cp .env.example .env   # then edit .env with your real key

python app.py

Open the local URL Gradio prints (usually http://127.0.0.1:7860), upload 2–3 PDFs, click Index Documents, then ask questions in the chat box.

---

RAG / personal_learning_rag — Personal Learning RAG

What it is: A fully local RAG pipeline over your own study materials (PDFs, markdown, text notes) — semantic search, grounded Q&A, and a daily "learning fact" pulled from a random passage in your library. No cloud API required: embeddings, the vector store, and generation all run on your machine via Ollama. See RAG/personal_learning_rag/README.md for the full pipeline, configuration options, and troubleshooting.

Stack: PyMuPDF + pytesseract + a local vision LLM (llava:7b) for PDF text extraction with OCR/vision fallback, langchain-text-splitters, sentence-transformers for embeddings, ChromaDB as the vector store, Ollama (qwen3:8b) for generation.

Run it:

cd RAG/personal_learning_rag
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

ollama pull qwen3:8b
ollama pull llava:7b

export MATERIALS_DIR="/path/to/your/materials"   # defaults to ~/Documents/Materials

python ingest.py      # one-time: chunk, embed, and store your materials
python search.py      # ask questions grounded in your materials
python daily_fact.py  # surface one random insight

---

CrewAI / agentic_rag_router — Agentic Router/Retriever RAG

What it is: A CrewAI-based multi-agent system that routes each question to the retrieval path best suited to answer it — a PDF vector search, a live web search (Tavily), or a direct LLM answer with no retrieval — via a Router Agent, then a Retriever Agent (scoped to exactly one tool) and an Answer Generation Agent. Every run's routing decision and per-agent reasoning is logged to outputs/reasoning_trace_log.csv. See CrewAI/agentic_rag_router/README.md for the full architecture, design rationale, and challenges.

Stack: CrewAI, crewai-tools (PDFSearchTool), LangChain community Tavily wrapper, OpenAI.

Run it:

cd CrewAI/agentic_rag_router
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

cp .env.example .env   # then edit .env with your real keys

jupyter notebook AgenticRAG_Project.ipynb

Run all cells top to bottom. The demo section runs 5 sample questions (PDF-routed, web-routed, direct, and out-of-scope) and prints each agent's reasoning, then saves the trace table to outputs/reasoning_trace_log.csv.

---

AutoGen / linkedin_automation — LinkedIn Content Automation

What it is: A FastAPI microservice that mimics an AutoGen/AG2 group-chat pipeline (Ideation → Drafting → Hashtags → Review/Scoring) for generating LinkedIn posts, orchestrated end-to-end by an n8n workflow that adds scheduling, a human approval gate, Slack review, and logging to Google Sheets. Full architecture diagram, error-handling notes, and a testing walkthrough are in AutoGen/linkedin_automation/README.md.

Stack: FastAPI, Pydantic, AG2 (AutoGen) for the real-LLM mode, n8n for orchestration.

Run it:

cd AutoGen/linkedin_automation
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

cp .env.example .env   # USE_MOCK=true works out of the box, no API key needed
uvicorn main:app --host 0.0.0.0 --port 8000 --reload

Test it:

curl -s http://localhost:8000/health
curl -s -X POST http://localhost:8000/linkedin \
  -H "Content-Type: application/json" \
  -d @sample_request.json | python3 -m json.tool

USE_MOCK=true (the default) runs fully offline deterministic agents — no OpenAI key required — so you can exercise the whole pipeline before switching to real AG2 agents with USE_MOCK=false.

The n8n workflow (n8n/FinEdge_LinkedIn_Automation.json) can be imported into a local n8n instance (npm install -g n8n && n8n start) to drive the microservice with scheduling, Slack approval, and logging — see the project's own README for the full setup and testing steps.

---

Notes

• Each project has its own .env.example next to its code — copy it to .env in that same folder and fill in real values. Each project's code loads its .env from its own folder explicitly, so it never reads config from another project or from a .env elsewhere in the repo.
• Each project's .env / API keys are excluded from git via .gitignore — never commit real API keys.
• venv/, __pycache__/, .DS_Store, and runtime-generated logs (*.log.jsonl) are also excluded.