rpanigrahi222/intruth-factcheck

1. What this repo is

- This is a Chrome Manifest V3 extension called *InTruth* / *realtime-factcheck*.

- Its purpose is to listen to the active tab’s audio, transcribe speech in real time, detect check-worthy claims, and show verdicts and evidence back to the user while the event is still happening.

2. Top-level runtime split

- The architecture is split across the standard extension contexts:

- Popup UI for user control and configuration

- Background service worker for orchestration

- Content script for in-page overlay and page interaction

- Offscreen document for audio/transcription work that can’t live in the service worker

- manifest.json is the key boundary definition for permissions, injections, and context wiring.

3. Popup as the control plane

- src/popup/popup.html, popup.js, and popup.css implement the extension’s front door.

- This is where the user likely starts/stops a session, supplies their bring-your-own Anthropic API key, and sets session options.

- The popup is the only obvious user-facing configuration surface; it likely writes to extension-local storage and sends commands into the runtime.

4. Background service worker as orchestrator

- src/background/service-worker-ex.js is the central coordinator.

- In an MV3 extension, this is typically responsible for:

- receiving commands from the popup,

- managing session state,

- triggering tab audio capture,

- routing messages between popup, offscreen page, and content script,

- and handling lifecycle/event transitions.

- This is the main integration boundary between UI intent and the real-time processing pipeline.

5. Offscreen document for audio processing

- src/offscreen/offscreen.html and offscreen-ex.js implement the audio-processing worker surface.

- The README and permissions indicate this is where tab audio capture / transcription happens, since the extension uses offscreen plus tabCapture.

- Architecturally, this isolates long-running media work from the ephemeral service worker and keeps transcription alive as a separate browser context.

6. Content script for on-page presentation

- src/content/overlay.js, overlay.css, and related files inject the in-page UI shown during a live session.

- This overlay is the user’s real-time view of detected claims, verdicts, and explanations.

- Because it runs in the page context, it acts as the bridge from extension state to something visually anchored on the active tab.

7. Claim and transcript feature extraction

- src/content/lexical-features.js suggests a local feature layer for analyzing text patterns, likely used to detect or rank candidate claims.

- src/content/session-export.js indicates session data can be packaged out of the live flow, probably for review or research use.

- These files suggest a separation between raw transcript handling and higher-level claim selection / session reporting.

8. Real-time data flow

- The likely flow is:

1. User starts a session in the popup.

2. Background service worker requests active-tab audio capture.

3. Offscreen page receives audio and performs transcription.

4. Transcript chunks are analyzed for check-worthy claims.

5. Claim text and context are sent for AI/evidence evaluation.

6. Results are returned to the background/content layer.

7. The overlay displays verdicts and sources in real time.

- This is a streaming pipeline, not a batch workflow.

9. External AI/service boundary

- The README states users provide their own Anthropic API key.

- That means the extension has a hard trust boundary: transcript data may leave the browser and be sent to the user-configured AI service for fact-checking.

- The repo explicitly positions this as local control of credentials, with no operator-managed API key backend.

10. Local state and privacy boundary

- The storage permission and README indicate user preferences and API configuration are stored locally in the browser.

- There is no sign of a server-side component in this repo.

- Architecturally, this is a client-only extension with external API calls only to user-chosen services.

11. Permissions-driven architecture

- The extension depends on tabCapture, activeTab, scripting, storage, and offscreen.

- These permissions directly map to the main subsystems:

- tabCapture → audio ingestion

- activeTab / scripting → page targeting and injection

- offscreen → media/transcription worker

- storage → local config/session state

12. Assets and packaging

- assets/ likely contains icons or static resources used by the extension UI and store packaging.

- LICENSE and README indicate a standalone open-source extension project, not a multi-service app.

13. Diagram takeaway

- The cleanest architecture diagram should show a browser extension with four contexts:

- user popup,

- background orchestrator,

- offscreen audio/transcription worker,

- content overlay injected into the page.

- Around that, show two external dependencies:

- the active browser tab’s audio/video stream as the input source,

- and Anthropic / other AI services as the verdict-generation boundary.