final changes

README.md

@@ -12,36 +12,70 @@ license: mit
 short_description: urn any Git commit into a human-readable engineering report.
 ---
 
-# CommitLens — AI-Powered Commit Analysis
+# CommitLens — AI-Powered Code Review Pipeline
 
-Analyses the latest commit of any GitHub repository and generates per-file summaries plus a final markdown report using [JetBrains Mellum 2](https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Thinking-GGUF-Q8_0).
+**CommitLens** is a high-performance information extraction and analysis pipeline that transforms raw GitHub diffs into structured, human-readable engineering reports. It uses a hybrid LLM approach: **JetBrains Mellum 2** for deep per-file analysis and **Groq-hosted Llama 3.3** for lightning-fast synthesis.
 
-## How it works
+![CommitLens UI](https://img.shields.io/badge/UI-Custom_HTML/CSS-blue)
+![Backend](https://img.shields.io/badge/Backend-Python_/_FastAPI-green)
+![LLM](https://img.shields.io/badge/LLM-Mellum_2_+_Groq_Llama_3.3-orange)
 
-1. **Enter a GitHub repo URL** (public or private)
-2. **Optionally provide a GitHub token** for higher rate limits or private repos
-3. **Click *Run Analysis*** — the app:
-   - Fetches the two most recent commits via the GitHub API
-   - Filters for source-code files (`.py`, `.js`, `.ts`, `.go`, `.rs`, etc.)
-   - Builds a structured prompt per file (commit metadata + before/after code + diff)
-   - Sends each prompt to **Mellum 2** for a concise per-file summary
-   - Combines all summaries and asks the model again for a comprehensive markdown report
-4. **View results** — per-file summaries and the final `.md` report side by side
+## 🚀 Key Features
 
-## Model
+- **Automated Diff Extraction**: Fetches the two latest commits from any GitHub repository and generates semantic diffs.
+- **Top-Impact Filtering**: Automatically identifies and prioritizes the most significant changes (top 2 files by lines changed) to ensure high-signal reviews.
+- **Hybrid LLM Pipeline**:
+  - **Mellum 2 (12B)**: Performs surgical, per-file code analysis. Optimized with 6-bit NF4 quantization for efficient GPU utilization.
+  - **Groq (Llama 3.3 70B)**: Generates a high-level executive summary and key takeaways in milliseconds.
+- **Cinematic UI**: A bespoke, low-latency frontend featuring a custom particle engine, real-time status tracking, and a "git-graph" hero visualization.
 
-This Space loads the **Mellum2-12B-A2.5B-Thinking-Q8_0** GGUF model (~7 GB) via `llama-cpp-python` on startup. First load may take several minutes.
+## 🛠 Tech Stack
 
-## Files
+- **Core**: Python 3.12, FastAPI, Gradio (Server Mode).
+- **ML/Inference**: `transformers`, `bitsandbytes` (4-bit/6-bit quantization), `torch`, `spaces` (ZeroGPU).
+- **APIs**: GitHub REST API, Groq Cloud API.
+- **Frontend**: Vanilla JavaScript (ES6+), HTML5 Canvas, CSS3 Grid/Flexbox.
+
+## 📂 Project Structure
 
 | File | Purpose |
 |------|---------|
-| `app.py` | Gradio application entry point |
-| `commitlens.py` | GitHub API pipeline (fetch, diff, filter, prompt builder) |
-| `requirements.txt` | Python dependencies |
+| `app.py` | Main application server; manages model lifecycle and GPU/API orchestration. |
+| `commitlens.py` | Data pipeline; handles GitHub API interaction, file filtering, and prompt engineering. |
+| `index.html` | Custom-built, high-fidelity frontend with interactive Git visualizations. |
+| `requirements.txt` | Dependency manifest (requests, gradio, torch, transformers, etc.). |
+
+## ⚙️ How It Works
+
+1. **Extraction**: The `GitHubClient` fetches commit metadata and raw patches.
+2. **Filtering**: Files are filtered by extension (keeping source code, ignoring binaries/locks) and sorted by impact.
+3. **Mellum Analysis**: The pipeline builds structured prompts containing "Before", "After", and "Diff" blocks. Mellum 2 generates concise summaries for each file.
+4. **Groq Synthesis**: Per-file summaries are batched and sent to Groq for a final structured Markdown report including a "Commit Overview" and "Key Takeaways".
+
+## 🛠 Setup & Usage
+
+### Local Development
+
+1. **Install dependencies**:
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+2. **Set Environment Variables**:
+   ```bash
+   export GROQ_API_KEY="your_groq_api_key"
+   ```
+
+3. **Run the application**:
+   ```bash
+   python app.py
+   ```
 
-## Requirements
+### CLI Mode
+You can also run the extraction pipeline directly:
+```bash
+python commitlens.py <github_repo_url> --token <optional_pat> --print-prompts
+```
 
-- `requests`
-- `gradio>=4.0.0`
-- `llama-cpp-python>=0.2.0`
+## 📄 License
+MIT

app.py

@@ -37,7 +37,7 @@ log = logging.getLogger("commitlens")
 # Config
 # ---------------------------------------------------------------------------
 
-MODEL_REPO_ID   = "JetBrains/Mellum2-12B-A2.5B-Thinking"
+MODEL_REPO_ID   = "JetBrains/Mellum2-12B-A2.5B-Instruct"
 GROQ_MODEL      = "llama-3.3-70b-versatile"   # fast Groq-hosted 70B
 # BATCH_TOKEN_BUDGET = 7000   # estimated input tokens; above this → sequential
 
@@ -202,9 +202,9 @@ def _generate_sequential(prompts: list[str]) -> list[str]:
         with torch.no_grad():
             out = _model.generate(
                 **enc,
-                max_new_tokens=16000,
+                max_new_tokens=1024,
                 use_cache=True,
-                do_sample=False,
+                do_sample=True,
                 temperature=0.4,
                 top_p=0.95,
                 pad_token_id=_tokenizer.pad_token_id,

commitlens.py

@@ -237,7 +237,7 @@ def fetch_commit_context(
         filtered, 
         key=lambda x: x.get("additions", 0) + x.get("deletions", 0), 
         reverse=True
-    )[:2]
+    )
 
     # Step 5 + 6 — build FileContext, fetch before/after content
     file_contexts: list[FileContext] = []
@@ -305,22 +305,22 @@ def build_prompts(ctx: CommitContext) -> list[str]:
         )
 
         # Before content
-        # if fc.before_content is not None:
-        #     sections.append(
-        #         "=== BEFORE CODE ===\n"
-        #         f"{fc.before_content}\n"
-        #     )
-        # else:
-        #     sections.append("=== BEFORE CODE ===\n(file did not exist)\n")
-
-        # # After content
-        # if fc.after_content is not None:
-        #     sections.append(
-        #         "=== AFTER CODE ===\n"
-        #         f"{fc.after_content}\n"
-        #     )
-        # else:
-        #     sections.append("=== AFTER CODE ===\n(file was deleted)\n")
+        if fc.before_content is not None:
+            sections.append(
+                "=== BEFORE CODE ===\n"
+                f"{fc.before_content}\n"
+            )
+        else:
+            sections.append("=== BEFORE CODE ===\n(file did not exist)\n")
+
+        # After content
+        if fc.after_content is not None:
+            sections.append(
+                "=== AFTER CODE ===\n"
+                f"{fc.after_content}\n"
+            )
+        else:
+            sections.append("=== AFTER CODE ===\n(file was deleted)\n")
 
         # Diff patch
         if fc.patch: