[JordanCoin]

Closes #59
Closes #57

#59: compact prompt for small models, thinking-model output

• _is_compact_model(model, setting) reads the parameter count from the tag. Models under 8B (mistral:7b, qwen2.5:1.5b, gemma3:270m) and the qwen3 family get the compact prompt; llama3.1:8b keeps the full prompt. MoE tags like 8x7b are skipped, a few known small tags (llama3.2, tinyllama, phi3) count as small, and unknown models get the full prompt. ai.compact_prompt: auto|true|false (env OPENFOIA_AI_COMPACT_PROMPT) overrides the rule. qwen3:8b on the full prompt returned truncated JSON even with think: false at 4000 and 8000 tokens.
• _build_extraction_prompt(..., compact=False) builds the compact form when asked: core types plus custom types, one line each, no relationship section, one short JSON example. The docstring now describes what the code does.
• _call_ollama and _call_openai strip <think>...</think> before returning. format: json is unchanged.

Local A/B, FOIA letter fixture, 14 expected entities:

model	full prompt	compact prompt
qwen3:8b	123s, truncated JSON, 0/14	27s, valid JSON, 12-13/14
llama3.1:8b	76s, 12/14	24-28s, 9-10/14

#57: junk mention scoring

These scores are now 0: anything over 80 characters, NER mentions under 3 characters, sentences (two sentence-ending periods, or a terminal period with 6+ words; U.S. and Dr. don't count), and NER mentions with 2+ garbled tokens.
These lower the score without dropping the mention: one garbled token (-0.30), a non-alphanumeric ratio above 0.25 (-0.50), PERSON over 5 tokens (-0.40).
The name-token check accepts any script, so José Müller passes.

Tests

tests/test_extraction_quality.py:

• the junk examples from the issue score 0 on every backend
• legitimate names score above 0.3, including U.S. Department of Justice and AT&T
• a table of model names and their size class
• the compact prompt has no relationship section and is under half the full prompt's length
• <think> stripping leaves parseable JSON

Full suite: 619 passed, 28 skipped. The 2 TestLLMExtraction failures are pre-existing: the default model llama3.2 is not pulled locally, so the LLM call fails before extraction.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FWQw8DDrDZ2tFzswSSEtce

[JordanCoin]

Review (Claude opus reviewer, read-only, 2026-11-06), two passes: MERGE. Pass 1 found the new junk scoring dropped 'X & Y' orgs (Kirkland & Ellis to 0.48) and six-word names ending in Inc. (hard zero), and penalized MONEY for punctuation; all fixed in the second commit and re-measured across a 35-name battery on regex and spaCy. Compact-prompt cutoff settled by A/B: qwen3:8b emits broken JSON on the full prompt at 4000 and 8000 tokens with thinking on or off, llama3.1:8b scores 12/14 on full, so auto = under 8B or the qwen3 family, overridable via ai.compact_prompt. Third commit (coordinator): six-plus-word mostly-lowercase mentions score 0 so a sentence ending in 'Inc.' no longer passes, and qwen3 is matched anywhere in the model name (hf.co/unsloth/Qwen3-8B-GGUF). Pre-existing and tracked separately: non-Latin names (東京都) and 2-letter names (Ng, 3M) score 0 because the letter count is ASCII-only.