Knowledge-Weaver 🕸️

Turning generic chat logs into a verified Source of Truth.

Knowledge-Weaver is a "Human-in-the-Loop" system that captures valuable information from team chats, verifies it, and makes it instantly searchable. It bridges the gap between ephemeral conversations and permanent organizational knowledge.

🎥 See it in Action

Watch the Full Demo Video

Architecture

The system uses a "Frankenstein" architecture, stitching together powerful components to create a seamless workflow:

graph TD
    subgraph "Frontend Layer"
        CE[Chrome Extension] -->|Capture & Search| API
        DB[HTML/JS Dashboard] -->|Manage & Verify| API
    end

    subgraph "Backend Layer"
        API[FastAPI Backend] -->|Process Requests| S[Services]
        S -->|Vector Search| VDB[ChromaDB]
        S -->|AI Analysis| LLM[Gemini 2.5 Flash]
    end

    subgraph "Data Layer"
        VDB <-->|Store Embeddings| DISK[Persist Directory]
        LLM -->|Context & Few-Shot| VDB
    end

    style CE fill:#f9f,stroke:#333,stroke-width:2px
    style DB fill:#bbf,stroke:#333,stroke-width:2px
    style API fill:#dfd,stroke:#333,stroke-width:2px
    style VDB fill:#fdd,stroke:#333,stroke-width:2px

Loading

The system is styled with a Custom Dark Theme (Spider Theme) to ensure a premium, consistent look across all components.

Core Concepts

🕷️ The Spider (Verification)

The "Spider" represents the verification layer. It ensures that only high-quality, human-reviewed information enters the permanent knowledge base.

• Source of Truth: Content marked as verified_human is treated as gold standard.
• Active Learning: The AI learns from these verified examples to improve future suggestions.

🕸️ The Weaver (Capture)

The "Weaver" is the capture mechanism (Chrome Extension & Auto-Ingestion).

• Draft Mode: Captured content starts as draft or verified_ai.
• Low Friction: Designed to be as effortless as possible to encourage high-volume capture.

🤖 The AI-Native Workflow

Knowledge-Weaver was built using a "Robot-First" philosophy. We treat AI Agents not just as coding assistants, but as First-Class Users of the application.

1. Evaluation > Labeling

Following modern AI engineering principles, we shifted our focus from manual data entry to Agent Evaluation. We use autonomous agents to run "Golden Path" regression tests on every commit, ensuring the system is operationally reliable.

2. Self-Healing Capabilities

By equipping our Dashboard with stable data-testid selectors, we enable AI agents (like the Antigravity Browser Subagent) to:

• Navigate the UI with 100% reliability.
• Identify Barriers: If an element is blocked, the agent logs the barrier.
• Self-Heal: The agent has the authority to fix broken selectors in the code to unblock itself.

3. Universal Design (Robot-First = Human-Ready)

We discovered that building for robots enforces strict semantic HTML and clear navigation paths. This "AI-Native" approach naturally results in a highly accessible application for human users relying on assistive technologies (Screen Readers), achieving a higher standard of Universal Design.

Key Features

1. Human-in-the-Loop Verification

AI suggests, Humans decide.

• Capture: AI suggests categories/tags.
• Review: Users verify/edit suggestions.
• Search: Verified content is boosted.

2. Active Learning (Dynamic Few-Shot Prompting)

The system gets smarter as you use it.

• Feedback Loop: When you edit a tag or category, the system records this as a "correction".
• Context Injection: Future analysis requests include these corrections as few-shot examples.
• Result: The AI mimics your team's specific style.

3. Gap-to-Gold

Identify and fill knowledge gaps.

• Tracking: Queries with zero results are tracked.
• Resolution: Experts can answer these gaps directly from the dashboard.

4. Recycle Bin (Soft Deletes)

• Safety: Deleted items are soft-deleted (hidden, not destroyed).
• Recovery: Restore items via the Recycle Bin tab.

5. Rich Content Editing

• Full Control: Edit summaries and content with a rich text interface.
• Seamless Updates: Changes are immediately reflected in the knowledge base.

6. Visual Source Verification

• Context Preservation: Captures screenshots of the source material (e.g., chat threads).
• Trust: Verify the origin of information with a visual reference.

7. Robot-Verified Reliability

• Accessibility: Validated by automated agents for DOM accessibility.
• Timezone Aware: Learning logs respect user's local timezone (UTC conversion).
• Auto-Redaction: Built-in HIPAA-compliant PII scrubbing.

Getting Started

Prerequisites

• Python 3.13+
• Chrome Browser

Installation

1. Clone the repository
2. Install dependencies: pip install -r requirements.txt
3. Set up environment variables (API Keys)

Running the System

1. Backend: uvicorn backend_api.main:app --reload
2. Dashboard: Open app_dashboard/index.html (Served via Python HTTP or FastAPI Static).
3. Extension: Load app_extension as an unpacked extension in Chrome.