feat: Enhanced adaptive speaker scraper with merged strategy

- Fixed model ID bug (strip openai/ prefix)
- Made max_tokens configurable for image extraction
- Enhanced screenshot scrolling to capture full pages
- Merged SmartScraperGraph + ScreenshotScraperGraph results
- Added hallucination filter for fake speakers
- Improved prompt to work with OpenAI content policies
- Added lazy-load scrolling support (timeout-based)
- Created FastAPI backend with web UI
- Added Excel export with metadata

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Author: 2 people

examples/ADAPTIVE_SCRAPER_README.md

@@ -0,0 +1,179 @@
+# 🎯 Adaptive Speaker Scraper
+
+Intelligent scraper that automatically detects website type and chooses the optimal scraping strategy.
+
+## 🧠 How It Works
+
+The scraper analyzes each website and classifies it into three types:
+
+### 1. **Pure HTML**
+- ✅ All speaker data in HTML text
+- 💰 **Strategy**: `SmartScraperGraph` (cheapest, fastest)
+- 📊 **Detection**: Completeness score ≥ 80%
+
+### 2. **Mixed Content**
+- ✅ Some data in HTML, some in images
+- 💰 **Strategy**: `OmniScraperGraph` (selective image processing)
+- 📊 **Detection**: 30-80% completeness + significant images
+- 🎯 Only processes relevant images (not all)
+
+### 3. **Pure Images**
+- ✅ All data embedded in images/widgets
+- 💰 **Strategy**: `ScreenshotScraperGraph` (full page screenshot)
+- 📊 **Detection**: Completeness score < 30% or no speakers found
+- 🎯 Sends 2 screenshots instead of 40+ individual images
+
+## 🚀 Usage
+
+### Basic Example
+
+```python
+from adaptive_speaker_scraper import scrape_with_optimal_strategy
+from pydantic import BaseModel, Field
+from typing import List
+
+class Speaker(BaseModel):
+    full_name: str = Field(default="")
+    company: str = Field(default="")
+    position: str = Field(default="")
+
+class SpeakerScrapeResult(BaseModel):
+    speakers: List[Speaker] = Field(default_factory=list)
+
+config = {
+    "llm": {
+        "api_key": "your-openai-key",
+        "model": "openai/gpt-4o-mini",
+    },
+    "verbose": True,
+}
+
+result = scrape_with_optimal_strategy(
+    url="https://example.com/speakers",
+    prompt="Extract all speakers with their names, companies, and positions",
+    config=config,
+    schema=SpeakerScrapeResult,
+)
+
+print(f"Strategy used: {result['strategy_used']}")
+print(f"Speakers found: {len(result['data']['speakers'])}")
+```
+
+### Run Demo
+
+```bash
+python examples/adaptive_speaker_scraper.py
+```
+
+## 🎛️ Decision Flow
+
+```
+Start
+  ↓
+Run SmartScraperGraph (fast, cheap)
+  ↓
+Analyze results:
+  - Completeness score
+  - Number of speakers
+  - Number of images
+  ↓
+┌─────────────────────┐
+│ Completeness ≥ 80%? │ → YES → ✅ Use SmartScraperGraph result
+└─────────────────────┘
+         ↓ NO
+┌─────────────────────────────────┐
+│ 30-80% complete + many images?  │ → YES → 🔄 Re-run with OmniScraperGraph
+└─────────────────────────────────┘
+         ↓ NO
+┌──────────────────────────────┐
+│ Very low data (<30%)?        │ → YES → 📸 Use ScreenshotScraperGraph
+└──────────────────────────────┘
+```
+
+## 💰 Cost Comparison
+
+### Example: 40 speakers on a page
+
+| Website Type | Strategy | API Calls | Cost (approx) |
+|-------------|----------|-----------|---------------|
+| Pure HTML | SmartScraperGraph | 1-2 text calls | $0.01 |
+| Mixed Content | OmniScraperGraph | 1 text + 20 images | $0.30 |
+| Pure Images | ScreenshotScraperGraph | 1 text + 2 screenshots | $0.05 |
+
+**Without adaptive detection**: Always using OmniScraperGraph with all images would cost **$0.50+**
+
+## 🔧 Customization
+
+### Adjust Detection Thresholds
+
+```python
+# In detect_website_type function:
+
+# More conservative (prefer cheaper strategies)
+if completeness >= 0.7:  # Lower from 0.8
+    website_type = WebsiteType.PURE_HTML
+
+# More aggressive image processing
+elif completeness >= 0.5:  # Higher from 0.3
+    website_type = WebsiteType.MIXED_CONTENT
+```
+
+### Control Image Processing
+
+```python
+# In scrape_with_optimal_strategy:
+omni_config["max_images"] = min(
+    analysis.get("num_images_detected", 10),
+    20  # Limit to 20 images maximum
+)
+```
+
+## 📊 Output Format
+
+```json
+{
+  "url": "https://example.com/speakers",
+  "website_type": "mixed_content",
+  "strategy_used": "OmniScraperGraph",
+  "analysis": {
+    "completeness_score": 0.45,
+    "num_speakers_found": 12,
+    "num_images_detected": 24
+  },
+  "data": {
+    "event": { ... },
+    "speakers": [ ... ]
+  }
+}
+```
+
+## 🎯 Best Practices
+
+1. **Start with gpt-4o-mini** for initial detection (cheap)
+2. **Upgrade to gpt-4o** if PURE_IMAGES detected (better vision)
+3. **Cache results** to avoid re-analyzing same URLs
+4. **Batch process** multiple URLs to optimize API usage
+
+## 🐛 Troubleshooting
+
+### "Not enough speakers extracted"
+- The page might be PURE_IMAGES but detected as MIXED_CONTENT
+- Solution: Lower the completeness threshold
+
+### "Too expensive"
+- Reduce `max_images` in OmniScraperGraph
+- Or force ScreenshotScraperGraph for image-heavy pages
+
+### "Missing some speakers"
+- Increase `max_images` for MIXED_CONTENT sites
+- Or use scroll/wait options in config for lazy-loaded content
+
+## 📚 Related Examples
+
+- `examples/frontend/batch_speaker_app.py` - Streamlit UI with manual strategy selection
+- `examples/smart_scraper_graph/` - Text-only extraction examples
+- `examples/omni_scraper_graph/` - Image+text extraction examples
+
+---
+
+**Key Advantage**: Automatically balances cost vs accuracy without manual intervention! 🎉