WikiTalk: Local Conversational Historian

<IN DEVELOPMENT - NOT CURRENTLY WORKING>

WikiTalk is an offline, conversational AI assistant that allows users to talk about history, science, and culture using a local copy of Wikipedia. It understands natural questions, supports follow-up questions and contextual discussion, and speaks answers aloud via Piper TTS.

Features

• Offline knowledge: Runs entirely from a local Wikipedia dataset (FineWiki Parquet files)
• Natural conversation: Multi-turn dialogue with context retention and topic continuity
• Voice interaction: Speaks answers using Piper TTS
• Grounded knowledge: Uses retrieved, cited chunks from Wikipedia to reduce hallucinations
• Mac-native: Optimized for Apple Silicon performance

Quick Start

1. Setup

# Install dependencies
python setup.py

# Or manually install requirements
pip install -r requirements.txt

2. Process Wikipedia Data

# This will create SQLite and FAISS indexes from the parquet files
python data_processor.py

Note: This step can take several hours depending on your system. The process will:

• Parse all parquet files in finewiki/data/enwiki/
• Create text chunks with overlap
• Build SQLite FTS5 index for BM25 search
• Generate embeddings and create FAISS index for dense retrieval

3. Start LLM Server

You need a local LLM server running. Options:

Option A: LM Studio

1. Download LM Studio from https://lmstudio.ai/
2. Load a model like Qwen2.5-14B-Instruct or Llama-3.1-8B-Instruct
3. Start the local server (usually on port 1234)

Option B: llama.cpp

# Download and build llama.cpp
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make

# Download a model (example with Qwen2.5-14B)
# Start server
./server -m models/qwen2.5-14b-instruct.gguf --port 1234

4. Optional: Setup TTS

For voice output, download Piper TTS voices:

# Create voices directory
mkdir -p voices

# Download a voice (example)
wget https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/en/en_US/amy/medium/en_US-amy-medium.onnx -O voices/en_US-amy-medium.onnx
wget https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/en/en_US/amy/medium/en_US-amy-medium.onnx.json -O voices/en_US-amy-medium.onnx.json

5. Run WikiTalk

python wikitalk.py

Usage

Once running, you can:

• Ask questions: "Tell me about the Meiji Restoration"
• Follow up: "And how did it affect Korea?"
• Get voice responses (if TTS is configured)
• Type clear to start a new conversation
• Type quit to exit

Architecture

User Input → Query Rewriter → Hybrid Retrieval → LLM → Response → TTS → Audio
                ↓
        Conversation Memory ← SQLite + FAISS Indexes

Components

• Data Processor: Parses parquet files and creates search indexes
• Hybrid Retriever: Combines BM25 (SQLite FTS5) and dense (FAISS) search
• LLM Client: Interfaces with local LLM for response generation
• TTS Client: Converts text to speech using Piper or macOS say
• Conversation Manager: Handles multi-turn dialogue context

Configuration

Edit config.py to customize:

• Data paths and model settings
• Retrieval parameters (top-k, chunk size)
• LLM server URL and model
• TTS voice settings

Requirements

• Python 3.8+
• 8GB+ RAM (for embeddings and FAISS index)
• 30GB+ disk space (for full English Wikipedia)
• Local LLM server (LM Studio or llama.cpp)
• Optional: Piper TTS for voice output

Troubleshooting

Data Processing Issues

• Ensure parquet files are in finewiki/data/enwiki/
• Check available disk space (30GB+ needed)
• Monitor memory usage during processing

LLM Connection Issues

• Verify LLM server is running on correct port
• Check LLM_URL in config.py
• Test with: curl http://localhost:1234/v1/models

TTS Issues

• Check Piper installation: which piper
• Verify voice files in voices/ directory
• Falls back to macOS say command if Piper unavailable

Performance

• Retrieval: <1 second for most queries
• Total response time: <10 seconds
• Memory usage: ~8GB for full English Wikipedia
• Storage: ~30GB for complete dataset

License

This project uses Wikipedia data under CC BY-SA 4.0 license.