Ask-the-Docs — Software Documentation Chatbot

CodeDocs AI

CodeDocs AI is a retrieval-augmented chatbot that answers questions about software documentation and source code. Simply provide a GitHub repository, documentation folder, or requirements specification, and the system automatically indexes the content to deliver accurate, source-cited answers.

🚀 Key Features

📄 Structure-Aware Document Ingestion

Instead of using fixed-size chunks for every document, the platform adapts chunking to each document type, ensuring retrieval happens at the most meaningful level.

Technical Documentation

• Markdown, MDX, reStructuredText, AsciiDoc
  • Split according to the heading hierarchy.
  • Code blocks and tables remain intact.

Requirements & Design Documents

• SRS, SDD, BRD (PDF, DOCX)
  • One chunk per numbered clause or requirement.
  • Requirements are tagged with identifiers such as FR-12 or REQ-004 for precise retrieval.

API Specifications

• OpenAPI, JSON, YAML
  • One chunk per API endpoint or structured record.

Source Code

• One chunk per function or class.
• Every chunk is tagged with its corresponding symbol name for accurate code-level retrieval.

General Documents

• HTML
• Jupyter Notebooks
• Plain Text

Automatically falls back to intelligent prose chunking when no structured format is detected.

🔍 Automatic Repository Discovery

Providing a GitHub repository URL automatically discovers and indexes relevant project documentation, including:

• README files
• docs/ directories
• User guides
• Technical specifications
• Architecture documents

Relevant files are scored and indexed without requiring manual selection.

🧠 Hybrid Retrieval Pipeline

• Dense semantic search using multilingual MiniLM embeddings (384 dimensions).
• BM25 keyword retrieval for lexical matching.
• Cross-encoder re-ranking to improve final retrieval accuracy.
• FAISS vector indexing for low-latency search.
• Per-tenant collections keep document indexes isolated between projects.

🎙 Voice & Multilingual Support

• Voice-based question input.
• Answers stream sentence-by-sentence for real-time text-to-speech.
• Barge-in detection automatically stops speaking when the user interrupts.
• Question language and response language can be configured independently.

---

🛠 Technology Stack

Layer	Technologies
Frontend	React, Vite
Backend	Python 3.11, FastAPI, LangChain
LLMs	Gemini, Groq
Vector Search	FAISS, MiniLM Embeddings, BM25, Cross-Encoder Re-ranking
Deployment	Google Cloud Run (CPU Always-On), Vercel
Testing	Pytest