Revert "fix: default all LLM backends to deterministic greedy decoding"

[phanindra-max]

Reverts #426 to fix pre-existing issues in the overall execution flows.

[github-actions]

Code Hygiene

The following items are defined but not yet in active use. Routine cleanup — no impact on functionality.

• laiser/llm_methods.py:87 — unused variable 'args' (100% confidence)
• laiser/llm_models/llama_cpp_handler.py:31 — unused variable 'chat_format' (100% confidence)
• laiser/llm_models/llama_cpp_handler.py:72 — unused variable 'exc' (100% confidence)
• laiser/llm_models/llama_cpp_handler.py:72 — unused variable 'exc_type' (100% confidence)
• laiser/llm_models/llama_cpp_handler.py:72 — unused variable 'tb' (100% confidence)
• laiser/llm_models/llm_router.py:70 — unused variable 'args' (100% confidence)
• laiser/llm_models/llm_router.py:85 — unused variable 'args' (100% confidence)
• laiser/services.py:130 — unused variable 'raw_description' (100% confidence)

[github-actions]

Complexity Analysis

Average: C — Moderate (14.678571428571429)

Hotspots tracked below. These are functional, well-exercised components — listed for visibility as the codebase scales.

Grade	Name	Type	Score
F	FAISSIndexManager.initialize_index	Method	51
E	SkillExtractionService.extract_and_align_core	Method	36
E	FAISSIndexManager.search_similar_skills	Method	32
D	ResponseParser._parse_skills_from_response	Method	24
D	FAISSIndexManager	Class	22
C	ResponseParser.parse_knowledge_task_response	Method	20
C	_BaseTaxonomyFAISSIndexManager.search_similar	Method	18
C	ResponseParser	Class	17
C	ResponseParser.parse_ksa_extraction_response	Method	16
C	ResponseParser.parse_skill_extraction_response	Method	15
C	openai_generate	Function	14
C	LLMRouter._initialize_components	Method	14
C	parse_output_vllm	Function	13
C	AlignmentService.align	Method	11
B	SkillExtractionService._deduplicate	Method	10
B	load_model_from_vllm	Function	10
B	_BaseTaxonomyFAISSIndexManager.initialize_index	Method	9
B	anthropic_generate	Function	9
B	ResponseParser.parse_ksa_details_response	Method	8
B	SkillExtractionService	Class	8
B	_BaseTaxonomyFAISSIndexManager	Class	8
B	GeminiAPI.generate_with_config	Method	8
B	SkillExtractorRefactored._expand_allowed_sources	Method	7
B	get_completion_vllm	Function	7
B	AlignmentService	Class	6
B	SkillExtractionService._normalize_mixed_concept_rows	Method	6
B	fetch_model_output	Function	6
B	LlamaCppBackend.__init__	Method	6

[github-actions]

📦 Preview package published!

Install and test this PR:

pip install --index-url https://test.pypi.org/simple/ laiser-preview==0.5.dev20260804025407

[Copilot]

🟡 Not ready to approve

It introduces at least one functional bug (Gemini max token limit parameters are ignored) and the paper now describes FAISS alignment as approximate/sub-linear despite the implementation using an exact IndexFlatIP scan.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

This review doesn't count toward merge requirements. Sign up for the private preview to control whether Copilot approvals count.

Pull request overview

Reverts the prior change that defaulted all LLM backends to deterministic greedy decoding, in order to restore the previous execution behavior across the extraction/alignment flow.

Changes:

• Removes deterministic decoding defaults (DEFAULT_TEMPERATURE, DEFAULT_TOP_P) and associated temperature/seed plumbing across router/service/extractor and backend helpers.
• Updates backend request/payload construction to no longer force deterministic decoding parameters (varies by backend).
• Removes the determinism-focused test module and updates project metadata/docs (pytest markers, paper text) accordingly.

File summaries

File	Description
tests/test_determinism.py	Removes determinism and cross-backend consistency test suite.
pytest.ini	Drops the determinism marker entry (keeps other markers).
paper.md	Updates narrative around backend coverage and alignment approach; currently mismatched with implementation details.
laiser/skill_extractor_refactored.py	Removes temperature/seed arguments and forwarding to the service layer.
laiser/services.py	Removes temperature/seed storage/forwarding into the router initialization path.
laiser/llm_models/openai.py	Removes temperature/top_p parameters from the public helper and payload construction.
laiser/llm_models/llm_router.py	Simplifies router init and stops forwarding temperature/seed into backend calls.
laiser/llm_models/llama_cpp_handler.py	Reverts deterministic defaults/seed handling; changes defaults and removes per-call seeding.
laiser/llm_models/hugging_face_llm.py	Removes temperature/seed-driven deterministic generation paths; simplifies sampling params.
laiser/llm_models/gemini.py	Removes temperature/seed parameters from the public API; retains fixed temperature config and exposes max token parameters (currently not fully applied without fixes).
laiser/llm_models/anthropic.py	Removes temperature/top_p parameters from the public helper and request payload.
laiser/llm_methods.py	Forces sampling on for HF generation calls and simplifies vLLM sampling params.
laiser/config.py	Removes deterministic decoding defaults while keeping GENERATION_SEED.

Review details

Suppressed comments (1)

laiser/llm_models/gemini.py:81

• gemini_generate() accepts max_output_tokens but doesn’t forward it into GeminiAPI, so the argument has no effect.

    client = GeminiAPI(api_key=api_key, model_name=model_name, timeout=timeout)

• Files reviewed: 13/13 changed files
• Comments generated: 3
• Review effort level: Lite

We're testing this review assessment. Please use 👍 or 👎 to tell us if it's correct.

[github-actions]

📦 Preview package published!

Install and test this PR:

pip install --index-url https://test.pypi.org/simple/ laiser-preview==0.5.dev20260804040738

[github-actions]

📦 Preview package published!

Install and test this PR:

pip install --index-url https://test.pypi.org/simple/ laiser-preview==0.5.dev20260804040833

[github-actions]

📦 Preview package published!

Install and test this PR:

pip install --index-url https://test.pypi.org/simple/ laiser-preview==0.5.dev20260804040953