IntelliCache

Intelligent semantic caching that reduces LLM latency, API costs, and redundant inference through response reuse.

---

Phase 1: AI Conversation Data Collector

Multi-Platform & Multi-Browser Data Collection Support

A unified, event-driven DOM observation and extraction extension for collecting real AI interaction datasets across multiple LLM web platforms (ChatGPT, Claude, and Gemini) and running reliably across major Chromium-based browsers (Google Chrome, Microsoft Edge, and Brave) and Mozilla Firefox.

Persists full query-response interactions, code blocks, and conversation metadata locally into IndexedDB through the repository layer.

Cross-Browser Compatibility Architecture

IntelliCache employs a clean, unified browser runtime abstraction (src/shared/browser.ts) that standardizes:

• Runtime Resolution: Auto-resolves browser.runtime (Firefox) and chrome.runtime (Chromium).
• Dual Message Semantics: Seamlessly handles both Promise-based messaging in Firefox and callback/sendResponse channel keeping in Chromium.
• Event-Driven Background Handling: Compatible with both Service Workers (Chromium MV3) and Event Pages (background.scripts in Firefox MV3).
• Extension Identity: Configured with Gecko ID (intellicache-collector@research.local) for stable IndexedDB namespace persistence.

---

Development & Build Commands

# Install dependencies
pnpm install

# Start Vite in development mode (with HMR)
pnpm dev

# Build the unpacked extension for Chromium (outputs to dist/)
pnpm build

# Build the unpacked extension for Mozilla Firefox (outputs to dist-firefox/)
pnpm build:firefox

# Build both Chromium and Firefox distribution targets
pnpm build:all

# Run Vitest unit & integration tests (28 test suites)
pnpm test

# Run TypeScript typecheck
pnpm exec tsc --noEmit

# Run ESLint
pnpm lint

# Check code formatting with Prettier
pnpm format:check

# Auto-format code with Prettier
pnpm format

---

Multi-Browser Loading Guide (Chrome, Edge, Brave, Firefox)

1. Google Chrome

1. Run pnpm build.
2. Navigate to chrome://extensions in the address bar.
3. Enable Developer mode using the toggle switch in the top-right corner.
4. Click Load unpacked in the top-left toolbar.
5. Select the dist/ directory inside this repository.

2. Microsoft Edge

1. Run pnpm build.
2. Navigate to edge://extensions in the address bar.
3. Enable Developer mode in the left sidebar.
4. Click Load unpacked at the top.
5. Select the dist/ directory.

3. Brave Browser

1. Run pnpm build.
2. Navigate to brave://extensions in the address bar.
3. Enable Developer mode toggle in the top-right corner.
4. Click Load unpacked.
5. Select the dist/ directory.

4. Mozilla Firefox

1. Run pnpm build:firefox.
2. Open Firefox and navigate to about:debugging#/runtime/this-firefox in the address bar.
3. Click Load Temporary Add-on....
4. Select manifest.json inside the dist-firefox/ directory (or any file within dist-firefox/).
5. The extension is now active with full ChatGPT, Claude, and Gemini collection support.

Verified Firefox Compatibility:

• Supported Versions: Firefox 109.0+ (Manifest V3 support), Firefox 115+ ESR, Firefox 128+ ESR, Firefox 130+ Current Release.
• Private Browsing Note: In Firefox Private Browsing mode, IndexedDB persistence is partitioned or restricted by Firefox by default. For reliable local collection, standard browsing mode is recommended.

IntelliCache/
├── src/
│   ├── background/
│   │   └── service-worker.ts       # Manifest V3 service worker & async DB coordinator
│   ├── content/
│   │   └── content.ts              # Content script entry point & adapter discovery
│   ├── database/
│   │   ├── db.ts                   # Dexie database class (IntelliCacheDB) & singleton
│   │   ├── index.ts                # Database module exports
│   │   ├── metrics.ts              # Character count, UTF-8 byte calculation utilities
│   │   ├── schema.ts               # Schema definitions and table indexes (v1)
│   │   ├── types.ts                # Interaction & Conversation entities, domain errors
│   │   └── repositories/
│   │       ├── conversation-repository.ts # Conversation data access & recency updates
│   │       └── interaction-repository.ts  # Interaction persistence & deduplication
│   ├── fingerprint/
│   │   ├── fingerprint.ts          # Web Crypto SHA-256 (3-tier fallback strategy)
│   │   ├── index.ts                # Fingerprint module exports
│   │   └── normalize.ts            # Isolated text normalization for hashing
│   ├── platforms/
│   │   ├── index.ts                # Platform module exports
│   │   ├── registry.ts             # Central adapter registry & URL-based discovery
│   │   ├── types.ts                # PlatformAdapter and ExtractedInteraction types
│   │   ├── chatgpt/
│   │   │   ├── adapter.ts          # MutationObserver, streaming stability & persistence
│   │   │   ├── parser.ts           # Pure extraction functions & code block preservation
│   │   │   └── selectors.ts        # ChatGPT semantic DOM selector definitions
│   │   ├── claude/
│   │   │   ├── adapter.ts          # Claude event-driven observer & persistence
│   │   │   ├── parser.ts           # Pure Claude extraction functions & code formatting
│   │   │   └── selectors.ts        # Claude semantic DOM selector definitions
│   │   └── gemini/
│   │       ├── adapter.ts          # Gemini event-driven observer & persistence
│   │       ├── parser.ts           # Pure Gemini extraction functions & code formatting
│   │       └── selectors.ts        # Gemini semantic DOM selector definitions
│   ├── popup/
│   │   ├── index.html              # Diagnostic popup UI
│   │   ├── popup.css               # Clean, accessible styling
│   │   └── popup.ts                # Popup controller & DB stats diagnostic
│   └── shared/
│       ├── types.ts                # Discriminated union messages & domain re-exports
│       └── messages.ts             # Message builders, type guards & URL platform detector
├── tests/
│   ├── chatgpt-adapter.test.ts     # ChatGPT adapter lifecycle, debouncing & DB integration
│   ├── chatgpt-fixtures.test.ts    # ChatGPT DOM snapshot fixtures (multi-turn, code, streaming)
│   ├── chatgpt-parser.test.ts      # Pure ChatGPT parser unit tests
│   ├── claude-adapter.test.ts      # Claude adapter lifecycle & DB integration tests
│   ├── claude-fixtures.test.ts     # Claude DOM snapshot fixtures (multi-turn, code, streaming)
│   ├── claude-parser.test.ts       # Pure Claude parser unit tests
│   ├── gemini-adapter.test.ts      # Gemini adapter lifecycle & DB integration tests
│   ├── gemini-fixtures.test.ts     # Gemini DOM snapshot fixtures (multi-turn, code, streaming)
│   ├── gemini-parser.test.ts       # Pure Gemini parser unit tests
│   ├── cross-platform.test.ts      # Cross-platform normalization contract tests
│   ├── database.test.ts            # Database initialization and lifecycle tests
│   ├── fingerprint.test.ts         # Deterministic fingerprinting & 3-tier strategy tests
│   ├── messages.test.ts            # Message contracts and URL detection tests
│   ├── metrics.test.ts             # UTF-8 byte counting and data model tests
│   ├── repositories.test.ts        # Interaction & Conversation repository CRUD tests
│   └── service-worker-db.test.ts   # Async message handling and DB error envelope tests
├── manifest.config.ts              # Type-safe Manifest V3 definition
├── vite.config.ts                  # Vite build setup with @crxjs/vite-plugin
├── vitest.config.ts                # Vitest test runner configuration
├── eslint.config.js                # Flat ESLint configuration
├── .prettierrc                     # Code style configuration
├── tsconfig.json                   # Strict TypeScript configuration
└── package.json                    # Scripts and dependencies

---

Normalized Data Contract

Every platform adapter (ChatGPTAdapter, ClaudeAdapter, GeminiAdapter) extracts conversation turns and produces the exact same normalized ExtractedInteraction contract:

Field	Type	Description
platform	'chatgpt' | 'claude' | 'gemini'	AI platform identifier
conversation_id	string | null	Namespaced conversation ID (chatgpt:${id}, claude:${id}, gemini:${id})
user_message_id	string | null	Platform user message ID when available (or null)
message_id	string | null	Platform assistant message ID when available (or null)
query.text	string	Extracted user prompt text
response.text	string	Extracted assistant response (with Markdown code blocks)
model.provider	string | null	Provider name (openai, claude, google)
model.name	string | null	Explicit model name if exposed by UI (or null)
conversation_title	string | null	Page/chat title with brand suffixes stripped
observed_at	string (ISO-8601)	Timestamp interaction was observed
capture_context	'on_load' | 'on_generate'	Provenance classification
source_timestamp	string | null	Platform timestamp if present in DOM
collector_version	string	Extension collector version (1.0.0)

---

Platform Adapter Implementations

1. ChatGPT (chatgpt.com, chat.openai.com)

   • User selector: div[class*="text-message"], div[class*="content"], [data-message-author-role="user"]
   • Assistant selector: .markdown, .prose, div[class*="markdown"]
   • Streaming guard: button[data-testid="stop-button"], .result-streaming, span.streaming-cursor
   • URL pattern: /c/{uuid}

2. Claude (claude.ai)

   • User selector: [data-testid="user-message"]
   • Assistant selector: [data-testid="assistant-message"], .font-claude-message
   • Streaming guard: button[aria-label="Stop Response"], [data-testid="streaming-indicator"], .streaming
   • URL pattern: /chat/{uuid}

3. Gemini (gemini.google.com)

   • User selector: user-query, [data-message-author-role="user"], .user-query-container
   • Assistant selector: model-response, [data-message-author-role="assistant"], .model-response-container
   • Streaming guard: button[aria-label="Stop response"], mat-spinner, .loading
   • URL pattern: /app/{id}, /chat/{id}

---

Privacy & Security Guarantee

• 100% Local: All interaction and conversation records are persisted strictly to the local browser IndexedDB.
• Zero Remote Telemetry: No tracking, external databases, analytics endpoints, or external network requests.
• No Credential Access: Extension never accesses session tokens, cookies, auth headers, or passwords.