LLM extraction: model-specific prompt templates + Qwen3 support

[JordanCoin]

Problem

Qwen3 8B produces excellent entities when given a simple prompt, but fails with our full extraction prompt. The prompt includes:

• 13 entity types (including obscure custom ones like SHELL_COMPANY_NAMES)
• 9 relationship types with Unicode arrows
• Detailed formatting instructions

Smaller models get confused by the complexity and return errors like "Please provide the full text."

What works

• Simple prompt: "Extract entities, return JSON with entities array" → perfect results
• Qwen3 8B correctly identifies Kirkland & Ellis, Clearview AI, Charlie Beck with high confidence
• The model normalizes "KIRKLAND 8. ELLIS LLP" → "Kirkland & Ellis LLP" (corrects OCR errors!)

Fix

1. Tiered prompts: Simple prompt for 3B-8B models, detailed prompt for 14B+
2. Strip custom types from prompt unless they have real patterns (most are placeholder junk)
3. Test with Qwen3 8B, llama3.1:8b, llama3.2:3b to ensure all work
4. Qwen3.5 (thinking models): Skip format: json constraint, strip <think> tags from response

Current state

• llama3.2:3b works (fast, lower quality)
• llama3.1:8b works (slow, higher quality)
• qwen3:8b works with simple prompts only
• qwen3.5:9b doesn't work (thinking model + JSON constraint incompatible)
• --model flag added to CLI
• Parser accepts "text"/"name"/"raw_text" keys from any model