community_contributions/dkisselev-zz/README.md

Digital Persona - Personal Knowledge Base

A RAG application that creates a queryable digital persona based on Facebook and LinkedIn data exports.

---

📁 Project Structure

dkisselev-zz/
├── persona_rag/                    # 🚀 Main Application
│   ├── persona_app.py             # Gradio chat interface
│   ├── ingest.py                  # Data ingestion + vector DB creation
│   ├── answer.py                  # RAG retrieval with configurable techniques
│   ├── evaluate.py                # Evaluation, tuning, RAG comparison
│   ├── tests.jsonl                # 25 test questions (LLM-generated) (.gitginored)
│   ├── pyproject.toml             # Dependencies (uv)
│   ├── README.md                  # 📘 Complete documentation
│   └── data/
│       ├── process_data.py        # Data processing (Facebook + LinkedIn)
│       ├── processed_facebook_data.json  # (.gitignored)
│       ├── processed_linkedin_data.json  # (.gitignored)
│       └── vector_db/             # Chroma database (.gitignored)
├── data_raw/
│   ├── facebook/                  # Raw Facebook export
│   └── linkedin/                  # Raw LinkedIn export
└── README.md                      # This file

---

📊 Data Overview

Data for Facebook and LinkedIn files is collected through data export functionality of the corresponding services and process usign process_data.py to tranform to json file that is locaed to ChromaDB for RAG

Facebook Data (Personal Life)

• Profile information (name, location, family, education)
• Years of posts and status updates
• Comments and social interactions
• Messages (privacy-preserving: only sent)
• Pages liked
• Events attended
• Group memberships
• Saved content
• Books read and app activities

LinkedIn Data (Professional Career)

• Professional profile and headline
• Career history
• Technical skills
• Professional certifications
• Education
• Colleague recommendations
• Projects
• Publications and thought leadership

Test Question Generation

The evaluation framework uses tests.jsonl - a collection of 25 test questions generated by an LLM based on your processed data.

Example test question:

{
  "question": "What is your current position?",
  "keywords": ["Tensor Lab", "Research Fellow", "current"],
  "reference_answer": "Research Fellow at The Tensor Lab, UCSF",
  "category": "career"
}

Generating tests:

1. Load your processed data into an LLM (GPT-4)
2. Prompting it to generate diverse test questions based on the content
3. Saving in JSONL format with required fields: question, keywords, reference_answer, category

Note: Questions could be generated manually as well without LLM participation

---

🎯 Architecture Highlights

Data Pipeline

Raw Data (Facebook + LinkedIn)
    ↓
Processing (first-person natural language)
    ↓
Grouping (semantic units by category & time)
    ↓
Chunking (1250 chars, 250 overlap)
    ↓
Embeddings (GTE-small, 384 dims)
    ↓
Vector DB (Chroma)

RAG Query Pipeline

User Query 
    ↓
(Query Expansion) → Sub-queries → 
Semantic Search → (Hybrid BM25) → 
Reranking → Context Retrieved
    ↓
LLM with RAG Context
    ↓
Agentic Tool Calls (as needed):
  • record_user_details (email capture)
  • record_unknown_question (improvement tracking)
  • push (Pushover notifications)
    ↓
Final Response

---

🎓 Key Features

Advanced RAG Techniques

• Query Expansion - Alternative phrasings for better coverage
• Hybrid Search - BM25 keyword + semantic search
• Sub-query Generation - Break complex questions into parts
• Cross-Encoder Reranking - Precision-focused ranking

Evaluation & Optimization

• Hyperparameter Tuning - Optimized chunk size
• RAG Comparison - Test all 4 configurations to find best approach
• Comprehensive Metrics - MRR, nDCG, Coverage, Accuracy, Completeness, Relevance
• LLM-as-Judge - Answer quality evaluation

Application Features

• Gradio Interface - Clean, interactive chat UI
• Agentic Architecture - Uses OpenAI function calling (tools) for intelligent actions
  • record_user_details - Captures email addresses and user information
  • record_unknown_question - Logs questions that cannot be answered for future improvement
  • push - Sends Pushover notifications for important interactions
• Smart Email Collection - Collects contact info once via tool calling, doesn't re-ask
• Conversation History - Multi-turn context management

---

🚀 Quick Start

# Navigate to application
cd persona_rag

# Install dependencies
uv sync

# Create .env file
echo "OPENAI_API_KEY=sk-proj-..." > .env

# Ingest data (if not already done)
uv run python ingest.py

# Launch application
uv run python persona_app.py

Open browser: http://127.0.0.1:7860

---

🛠️ Development Commands

# Data Processing
cd persona_rag/data
python process_data.py facebook          # Process Facebook only
python process_data.py linkedin          # Process LinkedIn only
python process_data.py both              # Process both

# Application
cd ..
uv run python persona_app.py             # Launch app

# Evaluation & Optimization
uv run python evaluate.py --compare-rag  # Compare RAG techniques
uv run python evaluate.py --tune         # Hyperparameter tuning
uv run python evaluate.py --eval         # Run evaluation
uv run python evaluate.py --all          # Everything

# RAG Technique Testing
uv run python evaluate.py --eval --query-expansion
uv run python evaluate.py --eval --hybrid-search