ollama backend never sets reasoning_effort, so local reasoning models (nemotron, deepseek-r1, qwq) can blow --api-timeout on real extraction chunks

[AutogenousEngine]

What happened

Benchmarking Claude/GLM-5.2/nemotron-3-super:120b-a12b-q4_K_M (via --backend ollama, served locally through Ollama) on identical corpora, Nemotron's extraction was consistently much sparser than the other two backends, and on a larger multi-file corpus one file (README.md) was dropped entirely after repeated timeout+bisect retries (5 files → 2 → 1, each hitting the client's 600s --api-timeout, with nothing left to bisect at 1 file).

No JSON parsing errors were involved — this is not the #2882 narration/parsing issue, which is already fixed and working correctly.

Root cause (confirmed empirically, not guessed)

Direct POST /v1/chat/completions against nemotron-3-super:120b-a12b-q4_K_M on the same Ollama host, same prompt (real source file + the extraction instructions), varying only reasoning_effort:

Call	content length	reasoning length	completion_tokens
default (reasoning_effort omitted)	3,299 chars	17,442 chars	5,234
reasoning_effort: "high"	2,092 chars	9,154 chars	2,807
reasoning_effort: "low"	3,463 chars	17,543 chars	5,244

Findings:

• choices[0].message has a separate reasoning field (Ollama's OpenAI-compat naming) alongside content. content itself is never empty or sparse in any of the three calls — each held valid, parseable JSON with genuine extraction — so this isn't the LLM responses not in JSON format: instructions instead of data #2882 recovery path, and it isn't a case of the answer hiding in the wrong field either.
• Ollama's OpenAI-compat endpoint does respect reasoning_effort for this model. "high" roughly halves both the reasoning length and total completion tokens compared to omitting it or passing "low".
• graphify/llm.py's BACKENDS["ollama"] config never sets a reasoning_effort key — only BACKENDS["gemini"] does ("reasoning_effort": "low"). So for every --backend ollama call, cfg.get("reasoning_effort") is None and the field is never sent, regardless of which model is actually being served.

Because the reasoning narration shares the same request's wall-clock and the client's --api-timeout (default 600s), a model that reasons at length before answering can burn most or all of that budget on narration alone. On a real extraction chunk (not the isolated single-file diagnostic above), that pushed a 5-file chunk into the adaptive bisect-and-retry path, and it still failed at a single file with no smaller unit to fall back to — the whole file was dropped from that run.

Why this isn't ollama-backend-wide

Ollama serves many non-reasoning models (the backend's own default is qwen2.5-coder:7b), so a blanket default would be wrong for most --backend ollama users. The bug is specifically that reasoning-tuned local models get no reasoning_effort treatment at all, whereas the same family is already recognized elsewhere in the codebase — _THINK_BLOCK_RE's own comment names nemotron, deepseek-r1, qwq as the models known to narrate a <think> block.

Repro

1. Serve a reasoning-tuned model locally, e.g. ollama pull nemotron-3-super:120b-a12b-q4_K_M.
2. graphify extract <a corpus with 3+ semantic files> --backend ollama --model nemotron-3-super:120b-a12b-q4_K_M --max-concurrency 1
3. On a large enough chunk, watch it bisect and potentially drop a file, with GRAPH_REPORT.md showing far fewer nodes/edges than the same corpus run through another backend (Gemini, GLM-5.2, Claude) gets.

Suggested fix

Opening a PR shortly: resolve reasoning_effort per-call for the ollama backend (env var override via GRAPHIFY_OLLAMA_REASONING_EFFORT, falling back to "high" for models matching the same reasoning-family list _THINK_BLOCK_RE already recognizes, and None/unchanged otherwise) rather than leaving it unset for every model served through Ollama.

[amtulifra]

hey, looked into this — Ollama never sets reasoning_effort, while the Gemini backend already does (llm.py:145), so extending the same treatment here, plus the GRAPHIFY_OLLAMA_REASONING_EFFORT override, seems straightforward. Mind assigning this to me so I can put up a PR?