AI providers: correlation decision and transcript correction move to OpenAI gpt-4o-mini

One funded account (OpenAI) now covers extraction, embeddings, summaries,
the ambiguous-scene correlation call and transcript correction; Groq stays
optional for STT. New app/internal/llm.py picks the provider from the model
name, so setting either model back to a gemini-* ID needs no code change.
Usage accounting for replay runs now covers both providers.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

drb-c2-core/app/config.py

@@ -133,12 +133,12 @@ class Settings(BaseSettings):
     # a failed LLM call falls back to the rules decision, so nothing broke
     # loudly. Verify against https://ai.google.dev/gemini-api/docs/models
     # before changing.
-    corr_cheap_model: str = "gemini-3.6-flash"   # was gemini-2.0-flash (shut down)
+    corr_cheap_model: str = "gpt-4o-mini"   # OpenAI since 2026-11-08 (was gemini-3.6-flash); a gemini-* ID still works
     # Transcript correction (server-26#36). Runs inside transcription, once per
     # transcribed call above MIN_WORDS_FOR_CORRECTION, so it is priced like STT
     # rather than like the correlation tier — cheap model on purpose.
     transcript_correction_enabled: bool = True
-    transcript_correction_model: str = "gemini-3.6-flash"
+    transcript_correction_model: str = "gpt-4o-mini"  # OpenAI since 2026-11-08; a gemini-* ID still works
     # Retry Whisper once when its output is degenerate. The same clip produced a
     # 56-word ten-code counting run on one attempt and real speech on the next
     # (2026-08-23, call e49ea32c), so a hallucination is a coin-flip rather than

drb-c2-core/app/internal/gemini.py

@@ -16,7 +16,6 @@ what it actually spent instead of an estimate.
 """
 import json
 import threading
-from contextvars import ContextVar
 from typing import Optional
 
 from app.config import settings
@@ -27,18 +26,6 @@ _client_lock = threading.Lock()
 # Models that rejected a thinking level: retried without one from then on.
 _no_thinking_level: set[str] = set()
 
-_usage_sink: ContextVar[Optional[dict]] = ContextVar("drb_gemini_usage", default=None)
-
-
-def collect_usage(sink: Optional[dict]):
-    """Route token counts for the current context into `sink` (a replay run). Returns a reset token."""
-    return _usage_sink.set(sink)
-
-
-def reset_usage(token) -> None:
-    _usage_sink.reset(token)
-
-
 def _get_client():
     global _client
     with _client_lock:
@@ -71,14 +58,8 @@ def _record(purpose: str, model: str, usage) -> None:
     prompt = getattr(usage, "prompt_token_count", None) or 0
     output = getattr(usage, "candidates_token_count", None) or 0
     thoughts = getattr(usage, "thoughts_token_count", None) or 0
-    logger.info(f"gemini usage {purpose} {model}: in={prompt} out={output} thinking={thoughts}")
-    sink = _usage_sink.get()
-    if sink is not None:
-        row = sink.setdefault(f"{purpose}:{model}", {"calls": 0, "in": 0, "out": 0, "thinking": 0})
-        row["calls"] += 1
-        row["in"] += prompt
-        row["out"] += output
-        row["thinking"] += thoughts
+    from app.internal.llm import record_usage
+    record_usage(purpose, model, prompt, output, thoughts)
 
 
 def generate_json(model: str, prompt: str, *, purpose: str,

drb-c2-core/app/internal/llm.py

@@ -0,0 +1,92 @@
+"""
+One entry point for the two small JSON-mode LLM jobs: the correlation decision
+on ambiguous scenes and transcript correction.
+
+The provider follows the model name: a `gemini-*` model goes through
+app/internal/gemini.py, anything else is an OpenAI chat model. Since the
+2026-11-08 provider consolidation both defaults are OpenAI (gpt-4o-mini), so
+one funded account (OpenAI) covers extraction, embeddings, summaries and these
+two jobs; Groq stays optional for transcription. Pointing either model setting
+back at a gemini-* ID needs no code change.
+
+Token counts from both providers land in the same usage sink, so a replay run
+(app/internal/replay.py) reports what it spent whichever provider served it.
+"""
+import json
+from contextvars import ContextVar
+from typing import Optional
+
+from app.config import settings
+from app.internal.logger import logger
+
+_usage_sink: ContextVar[Optional[dict]] = ContextVar("drb_llm_usage", default=None)
+
+
+def collect_usage(sink: Optional[dict]):
+    """Route token counts for the current context into `sink` (a replay run). Returns a reset token."""
+    return _usage_sink.set(sink)
+
+
+def reset_usage(token) -> None:
+    _usage_sink.reset(token)
+
+
+def record_usage(purpose: str, model: str, prompt: int, output: int, thinking: int = 0) -> None:
+    logger.info(f"llm usage {purpose} {model}: in={prompt} out={output} thinking={thinking}")
+    sink = _usage_sink.get()
+    if sink is not None:
+        row = sink.setdefault(f"{purpose}:{model}", {"calls": 0, "in": 0, "out": 0, "thinking": 0})
+        row["calls"] += 1
+        row["in"] += prompt
+        row["out"] += output
+        row["thinking"] += thinking
+
+
+def provider_for(model: str) -> str:
+    return "gemini" if (model or "").startswith("gemini") else "openai"
+
+
+def available(model: str) -> bool:
+    """Whether the provider serving `model` has a key configured."""
+    key = settings.gemini_api_key if provider_for(model) == "gemini" else settings.openai_api_key
+    return bool(key)
+
+
+def timeout_seconds(model: str) -> float:
+    if provider_for(model) == "gemini":
+        return settings.gemini_timeout_seconds
+    return (1 + max(0, settings.openai_max_retries)) * settings.openai_timeout_seconds
+
+
+def billing_fix(model: str) -> str:
+    if provider_for(model) == "gemini":
+        return "Top up billing at https://ai.studio/projects"
+    return "Top up billing at https://platform.openai.com/settings/organization/billing"
+
+
+def generate_json(model: str, prompt: str, *, purpose: str,
+                  thinking_level: Optional[str] = "minimal") -> dict:
+    """
+    Synchronous (run it via asyncio.to_thread). Returns the parsed JSON body.
+    Raises on API failure so callers' ai_health classification (billing /
+    dead model / transient) is unchanged. `thinking_level` only applies to
+    Gemini 3.x; OpenAI's gpt-4o family has no such setting.
+    """
+    if provider_for(model) == "gemini":
+        from app.internal import gemini
+        return gemini.generate_json(model, prompt, purpose=purpose, thinking_level=thinking_level)
+
+    from app.internal.ai_clients import openai_client
+    # JSON mode requires the word "JSON" in the prompt; both prompts say
+    # "Return JSON" / "Respond with JSON only".
+    resp = openai_client().chat.completions.create(
+        model=model,
+        messages=[{"role": "user", "content": prompt}],
+        response_format={"type": "json_object"},
+        temperature=0,
+    )
+    usage = getattr(resp, "usage", None)
+    record_usage(purpose, model,
+                 getattr(usage, "prompt_tokens", 0) or 0,
+                 getattr(usage, "completion_tokens", 0) or 0)
+    return json.loads(resp.choices[0].message.content or "{}")

drb-c2-core/app/internal/llm_correlator.py

@@ -1,5 +1,5 @@
 """
-LLM incident correlator (Gemini, settings.corr_cheap_model).
+LLM incident correlator (settings.corr_cheap_model, via app/internal/llm.py).
 
 One function is exposed:
   decide(call_id, ctx) — called by routers/upload.py ONLY when the rules
@@ -182,12 +182,12 @@ def _build_decide_prompt(ctx: dict, candidates: list[dict]) -> str:
 
 
 # ─────────────────────────────────────────────────────────────────────────────
-# Gemini API call (sync, runs in thread pool)
+# LLM call (sync, runs in thread pool)
 # ─────────────────────────────────────────────────────────────────────────────
 
 def _sync_gemini(model_name: str, prompt: str) -> dict:
-    from app.internal import gemini
-    return gemini.generate_json(model_name, prompt, purpose="correlation")
+    from app.internal import llm
+    return llm.generate_json(model_name, prompt, purpose="correlation")
 
 
 # ─────────────────────────────────────────────────────────────────────────────
@@ -253,12 +253,13 @@ def _parse_response(raw: dict, ctx: dict, candidates: list[dict]) -> Optional[di
 async def decide(call_id: str, ctx: dict) -> Optional[dict]:
     """
     One cheap-model correlation decision for an ambiguous scene. Returns None
-    (keep the rules decision) when GEMINI_API_KEY is unset, there is no
-    candidate to show, the answer fails validation, or Gemini fails.
+    (keep the rules decision) when the model's provider has no API key, there
+    is no candidate to show, the answer fails validation, or the call fails.
     """
     from app.internal.incident_correlator import llm_candidates
 
-    if not settings.gemini_api_key:
+    from app.internal import llm
+    if not llm.available(settings.corr_cheap_model):
         return None
     candidates = llm_candidates(ctx)
     if not candidates:
@@ -311,20 +312,23 @@ async def _log_llm_failure(where: str, tier: str, call_id: str, model: str, exc:
     # dead account never does. ai_health.classify() keeps a burst of rate limits
     # at WARNING while an empty account escalates like a bad model ID.
     if kind == "billing":
-        await _log_tier_down(where, tier, model, "the Gemini account is out of credit",
-                             "Top up billing at https://ai.studio/projects", text)
+        from app.internal import llm
+        await _log_tier_down(where, tier, model, f"the {llm.provider_for(model)} account is out of credit",
+                             llm.billing_fix(model), text)
         return
 
     logger.warning(f"{where} failed for call {call_id}: {text}")
+    from app.internal import llm
     await ai_health.report_degraded(
-        tier, "gemini", model, "transient API error",
+        tier, llm.provider_for(model), model, "transient API error",
         "no action needed unless this persists", permanent=False,
     )
 
 
 async def _log_tier_down(where: str, tier: str, model: str, problem: str, fix: str, text: str) -> None:
     """ERROR once per model, not once per call — this runs at radio-traffic volume."""
-    await ai_health.report_degraded(tier, "gemini", model, problem, fix, permanent=True)
+    from app.internal import llm
+    await ai_health.report_degraded(tier, llm.provider_for(model), model, problem, fix, permanent=True)
     if model in _dead_models:
         return
     _dead_models.add(model)

drb-c2-core/app/internal/recorrelation_sweep.py

@@ -70,13 +70,13 @@ def pipeline_worst_case_seconds(n_scenes: int = _BOUND_SCENES) -> float:
     from app.internal.transcription import stt_targets
     stt_providers = max(1, len(stt_targets()))
     stt = stt_providers * stt_rounds * openai_attempts * settings.stt_timeout_seconds
-    # Transcript correction (inside transcribe_call). gemini.generate_json only
-    # re-issues a request when the model rejects a thinking level — a fast
-    # 400, not a timeout — and google-genai itself does not retry.
-    correction = settings.gemini_timeout_seconds
+    # Transcript correction and the correlation decision go through
+    # llm.generate_json; the bound depends on which provider serves the model.
+    from app.internal import llm
+    correction = llm.timeout_seconds(settings.transcript_correction_model)
     extraction = openai_attempts * settings.openai_timeout_seconds
     embeddings = n_scenes * openai_attempts * settings.openai_timeout_seconds
-    correlation = settings.gemini_timeout_seconds  # at most one LLM call, scene 0 (#299)
+    correlation = llm.timeout_seconds(settings.corr_cheap_model)  # at most one LLM call, scene 0 (#299)
     return stt + correction + extraction + embeddings + correlation + _NON_AI_SLACK_SECONDS
 
 

drb-c2-core/app/internal/replay.py

@@ -474,9 +474,9 @@ async def _run(run_id: str, org_id: str, calls: list[dict], mode: str,
     fl_token = force_flags(_flags_for(mode))
     ai_failures: list = []
     ai_token = ai_health.collect_sandbox_failures(ai_failures)
-    from app.internal import gemini
+    from app.internal import llm
     usage: dict = {}
-    usage_token = gemini.collect_usage(usage)
+    usage_token = llm.collect_usage(usage)
     try:
         sem = asyncio.Semaphore(PREFETCH)
 
@@ -577,7 +577,7 @@ async def _run(run_id: str, org_id: str, calls: list[dict], mode: str,
         logger.error(f"Replay {run_id} failed: {e}")
     finally:
         ai_health._sandbox_failures.reset(ai_token)
-        gemini.reset_usage(usage_token)
+        llm.reset_usage(usage_token)
         unforce_flags(fl_token)
         fstore.exit_sandbox(sb_token)
         _cancel.discard(run_id)

drb-c2-core/app/internal/transcript_correction.py

@@ -253,11 +253,11 @@ def build_context_block(context: dict, talkgroup_name: Optional[str]) -> str:
 
 
 def _sync_gemini(model_name: str, prompt: str) -> dict:
-    from app.internal import gemini
+    from app.internal import llm
     # Correction rewrites text against vocabulary; keep a little reasoning
     # ("low") rather than the correlator's "minimal" until a replay shows
     # minimal doesn't hurt it.
-    return gemini.generate_json(model_name, prompt, purpose="correction", thinking_level="low")
+    return llm.generate_json(model_name, prompt, purpose="correction", thinking_level="low")
 
 
 async def correct(
@@ -282,7 +282,8 @@ async def correct(
     so a correction that only fixed the joined transcript would never reach the
     model on exactly the multi-transmission calls that carry the most content.
     """
-    if not settings.gemini_api_key or not settings.transcript_correction_enabled:
+    from app.internal import llm
+    if not llm.available(settings.transcript_correction_model) or not settings.transcript_correction_enabled:
         return None, None, False
 
     if len((text or "").split()) < MIN_WORDS_FOR_CORRECTION: