PageSense

PageSense is a browser-aware conversational AI Chrome Extension that enables users to ask intelligent questions about the content and structure of any webpage. Unlike traditional chatbots that operate on isolated text input, PageSense understands webpages in context by extracting page text, visual metadata (images, captions, infoboxes, and chart labels), and combining it with conversational reasoning.

The system runs entirely locally using a lightweight large language model via Ollama, streams responses token-by-token for real-time interaction, and maintains both short-term and long-term memory using a vector database. Delivered as a Chrome extension, PageSense acts as a private, efficient, and extensible AI assistant that reasons about what a page contains, not just what a user types.

Features:

1. Browser-Aware Understanding

• Extracts and analyzes webpage text
• Captures visual metadata including:
  • Images and alt text
  • Figure captions (including Wikipedia-specific structures)
  • Infobox content
  • Chart and graphic labels (SVG text)
• Enables reasoning on text-heavy and visual-rich pages

2. Conversational LLM (Local)

• Uses LLaMA 3.2 (1B Instruct) via Ollama
• Runs fully on CPU (no GPU required)
• Ensures local, private inference
• Produces coherent, context-aware responses

3. Real-Time Streaming Responses

• Token-by-token response streaming
• Immediate feedback in the Chrome extension UI
• Stop / cancel generation support

4. Memory System

• Short-term memory for conversational continuity
• Long-term semantic memory using Milvus (Zilliz Cloud)
• Retrieves relevant past interactions to enrich responses

5. Chrome Extension Interface

• Ask questions about the currently open webpage
• Clean, modern UI with gradient background
• Integrated page context extraction via content scripts
• Designed for low latency and smooth interaction

Roadmap

Stage 2 (Next)

• TTS/STT

Stage 3

• Screenshot-based visual reasoning
• Vision-language models (LLaVA / Qwen-VL)

Stage 4

• Multimodal embeddings
• Page-type auto-detection
• Cached visual summaries