Ollama/cloud chat-model parity: MoE + structured-output + no-think

Context

Sibling of #13853 (both sub of epic #12740). #13853 fixes the local (LM Studio / openAiCompatible) chat-model path — switch to the gemma-4-26b-a4b MoE, no-think, provider json_schema. The cloud/production path uses ollama (aiConfig.ollama.model, NEO_OLLAMA_MODEL, default gemma4:31b). The same root cause (dense 31B slow cold-prefill) likely applies on the cloud unless its GPU is materially faster — unverified. This ticket is the ollama-side parity, scoped to proceed in parallel with #13853.

Release classification: post-release (cloud REM pipeline; not release-blocking) → boardless.

The Problem

The ollama path differs from LM Studio in three ways, each requiring verification (NOT yet empirically confirmed — V-B-A required at implementation, unlike #13853 which was fully benchmarked):

1. Model tag — the MoE's ollama identifier differs from LM Studio's google/gemma-4-26b-a4b (ollama uses its own registry tags; today gemma4:31b). The MoE tag must be found / ollama pull-ed and set in NEO_OLLAMA_MODEL.
2. Structured output — ollama's native chat API uses a format: <JSON schema> field, not OpenAI-style response_format:{type:'json_schema'}. The ollama provider path must emit the caller schema in ollama's expected shape.
3. No-think toggle — ollama may not honor reasoning_effort; the gemma-MoE no-think mechanism under ollama may be a prompt token (/no_think) or a different param. Verify, then wire to the same per-task reasoningEffort config leaves introduced in #13853.

The Architectural Reality

• The ollama provider path (ai/provider/Ollama.mjs or the ollama branch of the provider dispatch — confirm at implementation).
• aiConfig.ollama.model leaf (default gemma4:31b).
• The cloud deployment runs ollama via compose-managed OLLAMA_* env (context length, parallel, keep-alive). Deployment specifics live in the private deployment repo — not here.
• Reuses #13853's per-task reasoningEffort config leaves (summaryReasoningEffort / graphReasoningEffort).

The Fix

1. Verify + set the ollama MoE model tag (NEO_OLLAMA_MODEL); document the ollama pull step.
2. Ollama provider emits structured output via ollama's native format field (with the caller's JSON schema).
3. Verify + wire the ollama no-think mechanism to the per-task reasoningEffort leaves.
4. Cloud benchmark — measure the ollama MoE cold-prefill vs the dense gemma4:31b on the cloud hardware; confirm the prefill win, or document if the cloud GPU makes the dense 31B acceptable (in which case the model-switch is local-only).

Contract Ledger Matrix

Decision Record impact

aligned-with ADR 0019 (config leaves resolved at use-site). Sibling of #13853; both realize epic #12740 (per-task model routing) for their respective providers.

Acceptance Criteria

• ollama MoE model tag identified, ollama pull-documented, set as NEO_OLLAMA_MODEL default (or env).
• ollama provider emits structured output via the native format field with the caller schema; extraction + summary produce valid schema-correct JSON.
• ollama no-think mechanism verified + wired to the per-task reasoningEffort leaves (#13853).
• Cloud benchmark: ollama MoE prefill measured vs dense gemma4:31b; result documented (win confirmed, or dense-31B-acceptable-on-cloud-GPU noted).

Out of Scope

• The local LM Studio / openAiCompatible path (#13853).
• The orchestrator ollama serving lifecycle (#13852 — ollama serve ensure + num_ctx preload); this ticket is model-choice + structured-output + no-think, not serving.

Avoided Traps

• Assuming ollama == LM Studio. ollama's structured-output (format) and no-think mechanisms are NOT the same as LM Studio's response_format:json_schema / reasoning_effort — each is a verify-then-wire task, not a copy of #13853.
• Assuming the cloud has the same prefill problem. Unverified — the cloud GPU may prefill the dense 31B acceptably; the benchmark decides whether the switch is cloud-needed.

Related

• Local sibling: #13853.
• Parent epic: #12740 (per-task model routing).
• Adjacent: #13852 (ollama serving lifecycle), #13851 (context verification).

Origin Session ID

2439c28b-245e-425a-85a7-ca7ee4aa0336

Handoff Retrieval Hints

• query_summaries: "ollama gemma MoE structured output format no-think cloud benchmark #13853"
• Live latest-open sweep: checked open issues at 2026-06-22T10:08Z; no equivalent (closest #13852 = serving lifecycle, not model-choice). Scoped for parallel pickup per operator direction.