LLM Context Extractor

A Chrome extension that extracts conversations from ChatGPT, Claude, and Gemini, then compresses them into structured context using the Gemini API — stripping noise while keeping decisions, goals, code, and unresolved questions.

---

What it does

LLM conversations get long. This extension captures a full conversation, runs it through a local scoring pipeline (chunker → scorer → extractor), and returns a compressed snapshot you can copy back into any LLM as clean context.

Raw conversation (100+ messages)
        ↓
  chunk + score each message
        ↓
  drop noise, compress mid, keep critical
        ↓
  Gemini extracts structured JSON
        ↓
{ goal, decisions, currentState, codeBlocks, unresolvedQuestions }

---

Features

• Extracts conversations directly from ChatGPT, Claude, and Gemini pages
• Compresses using a weighted scoring model (recency + type + keywords + cross-references)
• Structured output — goal, decisions, current state, unresolved questions, code blocks
• Accordion UI — each section collapsible; code blocks expand inline with syntax highlighting
• Copy to clipboard — copies context without raw code (replaces code with language + description)
• History tab — last 30 sessions saved locally, click any to reload

---

Project structure

GetllmContent/
├── manifest.json
├── background.js          # service worker — loads utils, handles messages
├── content.js             # runs on LLM pages, extracts raw messages
├── popup.html / popup.js  # extension UI
├── styles.css
└── utils/
    ├── parser.js          # page-specific DOM parsers (ChatGPT / Claude / Gemini)
    ├── modelTypes.js      # Message, ScoredChunk, CodeBlock, etc.
    ├── chunker.js         # tags each message with a ChunkType
    ├── scorer.js          # scores chunks: keep / compress / drop
    └── llm_extractor.js   # builds prompts, calls Gemini API, parses response

---

Setup

1. Clone the repo

git clone https://github.com/MENT0Z/context-compressor.git

2. Load into Chrome

• Go to chrome://extensions
• Enable Developer mode (top right)
• Click Load unpacked → select the GetllmContent folder

3. Add your Gemini API key

• Open the extension popup
• Paste your key into the API key field → click Save
• Get a key at aistudio.google.com

---

Usage

1. Open a ChatGPT, Claude, or Gemini conversation
2. Click the extension icon
3. Click ① Extract — grabs all messages from the page
4. Click ② Process — sends to Gemini, returns compressed context (10–30s)
5. Expand sections to read, or Copy to paste into a new chat

---

How scoring works

Each message gets a score from 0–1 based on four factors:

Factor	Weight	What it checks
Recency	35%	Last 20% of conversation scores highest
Type	30%	Code 0.95 · Decision 0.90 · Correction 0.85 · Info 0.50 · Question 0.40
Keywords	20%	Hits on: error, fix, schema, api, auth, final, etc.
Cross-reference	15%	Is this message referenced by later messages?

Scores ≥ 0.65 → keep verbatim · 0.35–0.65 → compress · < 0.35 → drop

---

Tech

• Chrome Extension Manifest V3
• Vanilla JS (no build step, no bundler)
• Gemini API (gemma-3-27b-it via REST)
• Token estimation: ~4 chars/token (no WASM dependency)

---

Limitations

• Parser accuracy depends on the LLM site's DOM — may break on UI updates
• Gemini API key is stored in chrome.storage.sync (not exposed to page scripts)
• No server, no account — everything runs locally except the Gemini API call