Chromium codebase FAQ Similarity Search Bot

A sophisticated AI-powered chatbot solution designed to help users quickly find answers to [Chromium codebase] questions. The bot uses vector embeddings and RAG (Retrieval Augmented Generation) to provide accurate, citation-backed responses.

Play with it In Production

This bot DOES NOT replace ChatGPT, Gemini, Copilot or other super-powerful models. Instead, it can quickly retrieve relevant information from your internal documents.

Essentially, this is a show case of what is possible with similarity search on your internal corporate data, while might not be best example.

🌟 Features

• Smart FAQ Matching: Uses vector embeddings for intelligent question matching
• RAG Fallback: Searches entire document corpus when FAQ doesn't match
• Citation-Based Responses: Every answer includes sources and confidence scores
• Static Deployment: Runs entirely in the browser - perfect for GitHub Pages
• Responsive Design: Works on desktop and mobile devices
• Real-time Processing: Fast in-memory vector search with cosine similarity

🏗️ Architecture

Phase 1: Context Database Preparation

• Vector embeddings of Q&A pairs using sentence transformers
• Fast in-memory vector index for similarity search
• Configurable similarity threshold for FAQ matching

Phase 2: User Query Processing

• FAQ matching with similarity threshold (default: 0.90)
• RAG fallback for document retrieval when FAQ confidence is low
• Citation-based responses with confidence scores

Phase 3: General Knowledge Retrieval

• Full document corpus search for unmatched queries
• Top-K retrieval with similarity ranking
• Synthesized answers with source attribution

📁 Project Structure

tactica.faq.similaritysearch/
├── src/
│   ├── TacTicA.FaqSimilaritySearchBot.Training/     
│   │   ├── Program.cs                      # C# training pipeline
│   │   ├── Services/                       # Vector embedding generation services
│   │   │   ├── SimpleEmbeddingService.cs      # Based on simplest hash-based vectors embeddings
│   │   │   ├── TfIdfEmbeddingService.cs       # Based on TF-IDF Vectorization and Keyword-Based Similarity
│   │   │   ├── OnnxEmbeddingService.cs        # Based on pre-trained Transformer Models (Customizable!)
│   │   │   └── DataProcessingService.cs       # Data processing & chunking
│   │   ├── models/                         # Input models files
│   │   │   ├── model_tokenizer.json           # https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2
│   │   │   ├── model_vocab.txt                # https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2
│   │   │   └── model.onnx                     # https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2
│   │   └── data/                           # Input data files
│   │       ├── questions_answers.txt          # Q&A pairs (you provide)
│   │       ├── wiki_links.txt                 # Wiki URLs (you provide)
│   │       └── exported-content/              # Exported content to work in full isolation
│   ├── TacTicA.FaqSimilaritySearchBot.Web/    # Static web application for simple hash-based similarity search
│   │   ├── index.html                              # Main UI
│   │   ├── css/style.css                           # Styling
│   │   └── js/chatbot.js                           # Client-side logic
│   ├── TacTicA.FaqSimilaritySearchBot.WebOnnx/  # Static web application for AI based search
│   │   ├── index.html                                  # Main UI
│   │   ├── css/style.css                               # Styling
│   │   └── js/chatbot.js                               # Client-side logic
│   └── TacTicA.FaqSimilaritySearchBot.Shared/  # Shared models and utilities
│       ├── Models/Models.css
│       └── Utils/VectorUtils.cs
├── wwwroot/data/                               # Generated embeddings & data
├── build.ps1                                   # Build & Training script
└── TacTicA.FaqSimilaritySearchBot.sln

🚀 Quick Start

Prerequisites

• .NET 9.0 SDK
• Python 3.x (optional, for local web server)
• Visual Studio Code or Visual Studio

1. Clone and Setup

git clone <your-repo-url>
cd tactica.faq.similaritysearch

2. Data

Data files in the TacTicA.FaqSimilaritySearchBot.Training/data/ directory:

data/questions_answers.txt (Q&A pairs, one per 2 lines):

Q: Where is android webview build instructions?
A: https://source.chromium.org/chromium/chromium/src/+/main:android_webview/docs/build-instructions.md

Q: What are frame trees in Chromium?
A: https://source.chromium.org/chromium/chromium/src/+/main:docs/frame_trees.md

data/wiki_links.txt (one URL per line):

https://source.chromium.org/chromium/chromium/src/+/main:docs/documentation_best_practices.md
https://source.chromium.org/chromium/chromium/src/+/main:docs/fuchsia/gtests.md

data/exported-content

Put as many as required exported content files in a simple markdown format into this directory. It allows system to work in full isolation without even reading wiki-links which is very helpful in case content is restricted via any type of authentication and you need to make some portion of data available to the bot without over-complicating things.

Format is following (URL is mandatory key field here):

URL: https://source.chromium.org/chromium/chromium/src/+/main:docs/fuchsia/gtests.md
# Content in Markdown starts here

Your actual content goes here...

3. Build and Train

# Build solution
dotnet restore
# Run training
dotnet build --configuration Release
# OR Build & Run together
cd "[PATH_TO_SOURCES]\src\TacTicA.FaqSimilaritySearchBot.Training" ; dotnet run --configuration Release

The process is super straightforward and NOT optimized. So it will run super slow... for couple of hours.

4. Serve Locally

Depending on the similarity search you want to use, make sure $WebPath in build.ps1 points to the right Web Project: TacTicA.FaqSimilaritySearchBot.WebOnnx - for AI based similarity search TacTicA.FaqSimilaritySearchBot.Web - for simplistic hash-based or TF-IDF similarity search

# Copy data to web server
Copy-Item -Path "[PATH_TO_SOURCES]\src\TacTicA.FaqSimilaritySearchBot.Training\wwwroot\data\*" -Destination "[PATH_TO_SOURCES]\src\TacTicA.FaqSimilaritySearchBot.WebOnnx\data\" -Recurse -Force
# Run
cd "[PATH_TO_SOURCES]\src\TacTicA.FaqSimilaritySearchBot.WebOnnx"; python -m http.server 8080

OR

# Copy data to web server
Copy-Item -Path "[PATH_TO_SOURCES]\src\TacTicA.FaqSimilaritySearchBot.Training\wwwroot\data\*" -Destination "[PATH_TO_SOURCES]\src\TacTicA.FaqSimilaritySearchBot.Web\data\" -Recurse -Force
# Run
cd "[PATH_TO_SOURCES]\src\TacTicA.FaqSimilaritySearchBot.Web"; python -m http.server 8080

🔧 Configuration

Training Configuration (src/TacTicA.FaqSimilaritySearchBot.Training/appsettings.json)

{
  "ProcessingSettings": {
    "MaxTokensPerChunk": 500,
    "SimilarityThreshold": 0.90,
    "TopKResults": 5
  }
}

Customizing Embeddings

The current implementation uses three possible embedding services for demo purposes: ── SimpleEmbeddingService - Vector embedding generation based on simplest hash-based vectors embeddings ── TfIdfEmbeddingService - Vector embedding generation based on TF-IDF Vectorization and Keyword-Based Similarity ── OnnxEmbeddingService - Vector embedding generation based on pre-trained sentence transformer model all-MiniLM-L6-v2 converted to ONNX format

📊 Data Flow

1. Training Phase:

   Q&A Text → Embeddings → Vector Index
   Wiki URLs → Content → Chunks → Embeddings → Document Index
   

2. Query Phase:

   User Question → Embedding → FAQ Search → High Score?
   ├─ Yes: FAQ Response + Related Docs
   └─ No:  RAG Search → Top Chunks → Generated Response
   

🔍 Known Issues

Low Accuracy Responses:

• We can increase similarity threshold in config
• We may add more Q&A pairs to training data
• Improve existing wiki content quality

Built with ❤️ AI