gemini: minimal thinking on correlation, token accounting per call

A day of replay runs (server-26#170) spent ~$5 of Gemini on ~7 two-hour
windows (~$0.70 per 290 calls) — several dollars a day per live deployment
for correlation alone — and nothing could say where it went (#45). Gemini
3.x thinks by default and bills it as output; the deprecated
google-generativeai SDK these calls used cannot set a thinking level.

- app/internal/gemini.py: every Gemini call (correlation + transcript
  correction) goes through google-genai with JSON mode, an explicit
  thinking level, and logs in/out/thinking tokens. A model that rejects
  the level is retried without it once and remembered, so the tier is
  never lost to a config param. API failures still raise for ai_health.
- correlator: thinking_level "minimal" (a link/new/orphan choice).
  transcript correction: "low" until a replay shows minimal is safe.
- replay: runs record real Gemini token usage (metrics.gemini_usage),
  shown in the Replay tab.
- requirements: google-genai.

c2-core: 474 pass. Frontend typecheck not run (no Node on this box).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Logan Cusano
2026-09-27 01:10:40 -04:00
co-authored by Claude Opus 5.5
parent 266c958208
commit 0543526eb0
8 changed files with 186 additions and 19 deletions
+94
View File
@@ -0,0 +1,94 @@
"""
One place every Gemini call goes through: JSON-mode generation, an explicit
thinking level, and token accounting.
Why it exists: a day of replay runs (server-26#170) cost ~$5 of Gemini for
~7 two-hour windows — roughly $0.70 per 290 calls, which projects to several
dollars a day per live deployment for correlation alone — and nothing in DRB
could say where it went (server-26#45). Gemini 3.x models "think" by default
and bill that as output; the old google-generativeai SDK these calls used
cannot even set a thinking level. A link/new/orphan choice or a transcript
cleanup does not need extended reasoning.
Every call logs its token counts, and inside a replay run they are also added
to the run's own usage sink (see app/internal/replay.py), so a run reports
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
from app.internal.logger import logger
_client = None
_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:
if _client is None:
from google import genai # lazy — only when a Gemini call is made
_client = genai.Client(api_key=settings.gemini_api_key)
return _client
def _config(thinking_level: Optional[str]):
from google.genai import types
kwargs = {"response_mime_type": "application/json"}
if thinking_level:
kwargs["thinking_config"] = types.ThinkingConfig(thinking_level=thinking_level)
return types.GenerateContentConfig(**kwargs)
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
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, exactly like the old per-module helpers, so callers'
ai_health classification (billing / dead model / transient) is unchanged.
"""
client = _get_client()
level = None if model in _no_thinking_level else thinking_level
try:
resp = client.models.generate_content(model=model, contents=prompt, config=_config(level))
except Exception as e:
# A model that doesn't accept this thinking level answers 400 for
# every call; drop the setting for that model rather than lose the tier.
if level and "thinking" in str(e).lower():
logger.warning(f"gemini: {model} rejected thinking_level={level!r} ({e}); retrying without it")
_no_thinking_level.add(model)
resp = client.models.generate_content(model=model, contents=prompt, config=_config(None))
else:
raise
_record(purpose, model, getattr(resp, "usage_metadata", None))
return json.loads(resp.text)
+2 -10
View File
@@ -20,7 +20,6 @@ Error handling: any Gemini failure returns None from decide() and the
rules_decision from tiebreak() so the pipeline never stalls. rules_decision from tiebreak() so the pipeline never stalls.
""" """
import asyncio import asyncio
import json
from datetime import datetime, timezone from datetime import datetime, timezone
from typing import Optional from typing import Optional
from app.internal.logger import logger from app.internal.logger import logger
@@ -190,15 +189,8 @@ def _build_tiebreak_prompt(rules_decision: dict, llm_decision: dict, ctx: dict)
# ───────────────────────────────────────────────────────────────────────────── # ─────────────────────────────────────────────────────────────────────────────
def _sync_gemini(model_name: str, prompt: str) -> dict: def _sync_gemini(model_name: str, prompt: str) -> dict:
import google.generativeai as genai # lazy import — only when needed from app.internal import gemini
return gemini.generate_json(model_name, prompt, purpose="correlation")
genai.configure(api_key=settings.gemini_api_key)
model = genai.GenerativeModel(
model_name,
generation_config={"response_mime_type": "application/json"},
)
response = model.generate_content(prompt)
return json.loads(response.text)
# ───────────────────────────────────────────────────────────────────────────── # ─────────────────────────────────────────────────────────────────────────────
+5
View File
@@ -469,6 +469,9 @@ async def _run(run_id: str, org_id: str, calls: list[dict], mode: str,
fl_token = force_flags(_flags_for(mode)) fl_token = force_flags(_flags_for(mode))
ai_failures: list = [] ai_failures: list = []
ai_token = ai_health.collect_sandbox_failures(ai_failures) ai_token = ai_health.collect_sandbox_failures(ai_failures)
from app.internal import gemini
usage: dict = {}
usage_token = gemini.collect_usage(usage)
try: try:
sem = asyncio.Semaphore(PREFETCH) sem = asyncio.Semaphore(PREFETCH)
@@ -561,6 +564,7 @@ async def _run(run_id: str, org_id: str, calls: list[dict], mode: str,
metrics = compute_metrics(incidents, sb_calls) metrics = compute_metrics(incidents, sb_calls)
metrics["est_cost_usd"] = _running_cost(progress, metrics, mode) metrics["est_cost_usd"] = _running_cost(progress, metrics, mode)
metrics["ai_failures"] = dict(Counter(f"{f['tier']}: {f['problem']}" for f in ai_failures)) metrics["ai_failures"] = dict(Counter(f"{f['tier']}: {f['problem']}" for f in ai_failures))
metrics["gemini_usage"] = usage
except Exception as e: except Exception as e:
status = "failed" status = "failed"
errors.append(f"run: {type(e).__name__}: {e}"[:300]) errors.append(f"run: {type(e).__name__}: {e}"[:300])
@@ -568,6 +572,7 @@ async def _run(run_id: str, org_id: str, calls: list[dict], mode: str,
logger.error(f"Replay {run_id} failed: {e}") logger.error(f"Replay {run_id} failed: {e}")
finally: finally:
ai_health._sandbox_failures.reset(ai_token) ai_health._sandbox_failures.reset(ai_token)
gemini.reset_usage(usage_token)
unforce_flags(fl_token) unforce_flags(fl_token)
fstore.exit_sandbox(sb_token) fstore.exit_sandbox(sb_token)
_cancel.discard(run_id) _cancel.discard(run_id)
@@ -43,7 +43,6 @@ another equally plausible word.
""" """
import asyncio import asyncio
import json
import re import re
from typing import Any, Optional from typing import Any, Optional
@@ -228,14 +227,11 @@ def build_context_block(context: dict, talkgroup_name: Optional[str]) -> str:
def _sync_gemini(model_name: str, prompt: str) -> dict: def _sync_gemini(model_name: str, prompt: str) -> dict:
import google.generativeai as genai # lazy import — only when needed from app.internal import gemini
# Correction rewrites text against vocabulary; keep a little reasoning
genai.configure(api_key=settings.gemini_api_key) # ("low") rather than the correlator's "minimal" until a replay shows
model = genai.GenerativeModel( # minimal doesn't hurt it.
model_name, return gemini.generate_json(model_name, prompt, purpose="correction", thinking_level="low")
generation_config={"response_mime_type": "application/json"},
)
return json.loads(model.generate_content(prompt).text)
async def correct( async def correct(
+1
View File
@@ -6,6 +6,7 @@ firebase-admin
google-cloud-storage google-cloud-storage
openai openai
google-generativeai google-generativeai
google-genai
numpy numpy
httpx httpx
python-multipart python-multipart
+70
View File
@@ -0,0 +1,70 @@
"""
app/internal/gemini.py — thinking level, fallback when a model rejects it,
and token accounting into a replay's usage sink (server-26#170 cost finding).
"""
from types import SimpleNamespace
from unittest.mock import patch
from app.internal import gemini
class _FakeModels:
def __init__(self, reject_thinking=False):
self.reject_thinking = reject_thinking
self.configs = []
def generate_content(self, model, contents, config):
self.configs.append(config)
if self.reject_thinking and config.get("thinking_level"):
raise RuntimeError("400 INVALID_ARGUMENT: thinking_level is not supported for this model")
return SimpleNamespace(
text='{"action": "link"}',
usage_metadata=SimpleNamespace(prompt_token_count=1200, candidates_token_count=30,
thoughts_token_count=0),
)
def _patched(models):
client = SimpleNamespace(models=models)
return (patch.object(gemini, "_get_client", return_value=client),
patch.object(gemini, "_config", lambda level: {"thinking_level": level}))
def test_minimal_thinking_by_default_and_usage_lands_in_the_sink():
models = _FakeModels()
a, b = _patched(models)
sink = {}
tok = gemini.collect_usage(sink)
try:
with a, b:
assert gemini.generate_json("m1", "p", purpose="correlation") == {"action": "link"}
finally:
gemini.reset_usage(tok)
assert models.configs == [{"thinking_level": "minimal"}]
assert sink == {"correlation:m1": {"calls": 1, "in": 1200, "out": 30, "thinking": 0}}
def test_model_that_rejects_thinking_level_falls_back_once():
models = _FakeModels(reject_thinking=True)
a, b = _patched(models)
gemini._no_thinking_level.discard("m2")
with a, b:
gemini.generate_json("m2", "p", purpose="correlation")
gemini.generate_json("m2", "p", purpose="correlation")
# first call: tried minimal, retried without; second call: straight without
assert models.configs == [{"thinking_level": "minimal"}, {"thinking_level": None}, {"thinking_level": None}]
gemini._no_thinking_level.discard("m2")
def test_other_failures_still_raise_for_ai_health():
class Boom(_FakeModels):
def generate_content(self, **kw):
raise RuntimeError("429 insufficient_quota")
a, b = _patched(Boom())
with a, b:
try:
gemini.generate_json("m3", "p", purpose="correlation")
except RuntimeError as e:
assert "insufficient_quota" in str(e)
else:
raise AssertionError("should raise")
@@ -361,6 +361,14 @@ function RunDetail({ run }: { run: ReplayRun }) {
paths: {Object.entries(m.corr_path).map(([k, v]) => `${k} ${v}`).join(" · ")} paths: {Object.entries(m.corr_path).map(([k, v]) => `${k} ${v}`).join(" · ")}
</p> </p>
)} )}
{m?.gemini_usage && Object.keys(m.gemini_usage).length > 0 && (
<p className="text-xs font-mono text-gray-500">
Gemini tokens:{" "}
{Object.entries(m.gemini_usage)
.map(([k, u]) => `${k} ${u.calls} calls, in ${u.in.toLocaleString()} / out ${u.out.toLocaleString()} / thinking ${u.thinking.toLocaleString()}`)
.join(" · ")}
</p>
)}
{m?.ai_failures && Object.keys(m.ai_failures).length > 0 && ( {m?.ai_failures && Object.keys(m.ai_failures).length > 0 && (
<p className="text-xs font-mono text-amber-400"> <p className="text-xs font-mono text-amber-400">
AI failures: {Object.entries(m.ai_failures).map(([k, v]) => `${k} ×${v}`).join(" · ")} AI failures: {Object.entries(m.ai_failures).map(([k, v]) => `${k} ×${v}`).join(" · ")}
+1
View File
@@ -340,6 +340,7 @@ export interface ReplayMetrics {
llm_decisions: number; llm_decisions: number;
est_cost_usd: number; est_cost_usd: number;
ai_failures?: Record<string, number>; ai_failures?: Record<string, number>;
gemini_usage?: Record<string, { calls: number; in: number; out: number; thinking: number }>;
} }
export interface ReplayRun { export interface ReplayRun {