Skip to content

:hedgehog: PostHog is the leading platform for building self-driving products. Our developer tools – AI observability, analytics, session replay, flags, experiments, error tracking, logs, and more – capture all the context agents need to diagnose problems, uncover opportunities, and ship fixes. Steer it all from Slack, web, desktop, or the MCP.

feat(growth): route ai enrichment labelling through the ai-gateway (#123407)

Jonathan Mieloo6309612ab2ec
diff --git a/products/growth/backend/api/ai_enrichment.py b/products/growth/backend/api/ai_enrichment.py
index 92801db8eff9..4ac2e6e995cf 100644
--- a/products/growth/backend/api/ai_enrichment.py
+++ b/products/growth/backend/api/ai_enrichment.py
@@ -35,7 +35,7 @@ from posthog.api.utils import ErrorResponseSerializer
from posthog.exceptions import Conflict
from posthog.exceptions_capture import capture_exception
from posthog.helpers.impersonation import is_impersonated
-from posthog.llm.gateway_client import get_llm_client
+from posthog.llm.gateway_client import build_openai_client
from posthog.models import User
from posthog.permissions import IsStaffUser
from posthog.renderers import SafeJSONRenderer
@@ -58,6 +58,7 @@ from products.growth.backend.enrichment.lab import (
stream_run_classifications,
)
from products.growth.backend.enrichment.labels import (
+ AI_ENRICHMENT_FEATURE,
UNKNOWN,
recent_latest_fetches_qs,
signup_domain_for_organization,
@@ -237,7 +238,11 @@ class AIEnrichmentViewSet(viewsets.ViewSet):
# stop_after_attempt(3)); the SDK's own internal retries underneath would multiply that
# budget several-fold per fetch and actively worsen a 429 the tenacity layer is already
# backing off from.
- client = get_llm_client(product="growth").with_options(max_retries=0)
+ client = build_openai_client(
+ product="growth",
+ ai_product="growth",
+ properties={"ai_feature": AI_ENRICHMENT_FEATURE, "surface": "staff_run"},
+ ).with_options(max_retries=0)
logger.info(
"growth_ai_enrichment_run",
diff --git a/products/growth/backend/api/test_ai_enrichment.py b/products/growth/backend/api/test_ai_enrichment.py
index 99bc4d84a2b4..f673468850d5 100644
--- a/products/growth/backend/api/test_ai_enrichment.py
+++ b/products/growth/backend/api/test_ai_enrichment.py
@@ -40,7 +40,7 @@ def _mock_llm_client(
verdict: bool = True, confidence: float = 0.9, reasoning: str = "builds ai software", verdict_key: str = "is_ai"
) -> MagicMock:
client = MagicMock()
- # The run action chains .with_options(max_retries=0) onto get_llm_client(...); without this,
+ # The run action chains .with_options(max_retries=0) onto build_openai_client(...); without this,
# that call returns a fresh, unconfigured child mock and every assertion below silently
# checks the wrong object (see enrichment_label_batch.py for the same pattern).
client.with_options.return_value = client
@@ -188,7 +188,7 @@ class TestAIEnrichmentAPI(APIBaseTest):
fake_client = MagicMock()
fake_client.models.list.return_value = [MagicMock(id="gpt-5.2"), MagicMock(id="claude-fable-5")]
- with patch("products.growth.backend.enrichment.lab.get_llm_client", return_value=fake_client):
+ with patch("products.growth.backend.enrichment.lab.build_openai_client", return_value=fake_client):
response = self.client.get("/api/growth_ai_enrichment/models/")
self.assertEqual(response.status_code, status.HTTP_200_OK)
@@ -196,7 +196,9 @@ class TestAIEnrichmentAPI(APIBaseTest):
self.assertEqual(ids, {"gpt-5.2", "claude-fable-5"})
def test_models_returns_empty_rather_than_a_stale_list_when_the_gateway_is_down(self):
- with patch("products.growth.backend.enrichment.lab.get_llm_client", side_effect=RuntimeError("gateway down")):
+ with patch(
+ "products.growth.backend.enrichment.lab.build_openai_client", side_effect=RuntimeError("gateway down")
+ ):
response = self.client.get("/api/growth_ai_enrichment/models/")
# No curated mirror: a hand-maintained fallback goes stale silently.
@@ -422,7 +424,7 @@ class TestAIEnrichmentAPI(APIBaseTest):
client = _mock_llm_client()
with (
- patch("products.growth.backend.api.ai_enrichment.get_llm_client", return_value=client),
+ patch("products.growth.backend.api.ai_enrichment.build_openai_client", return_value=client),
patch("products.growth.backend.enrichment.lab.ai_processing_approved", return_value=False),
):
response = self.client.post(
@@ -513,7 +515,7 @@ class TestAIEnrichmentAPI(APIBaseTest):
)
client = _mock_llm_client()
- with patch("products.growth.backend.api.ai_enrichment.get_llm_client", return_value=client) as get_client:
+ with patch("products.growth.backend.api.ai_enrichment.build_openai_client", return_value=client) as get_client:
response = self.client.post(
"/api/growth_ai_enrichment/run/",
{
@@ -528,7 +530,11 @@ class TestAIEnrichmentAPI(APIBaseTest):
)
_drain_ndjson(response.streaming_content) # type: ignore[attr-defined]
- get_client.assert_called_once_with(product="growth")
+ get_client.assert_called_once_with(
+ product="growth",
+ ai_product="growth",
+ properties={"ai_feature": "ai_enrichment", "surface": "staff_run"},
+ )
client.with_options.assert_called_once_with(max_retries=0)
def test_run_is_throttled_for_session_auth(self):
@@ -547,7 +553,7 @@ class TestAIEnrichmentAPI(APIBaseTest):
"sample": 1,
}
- with patch("products.growth.backend.api.ai_enrichment.get_llm_client", return_value=_mock_llm_client()):
+ with patch("products.growth.backend.api.ai_enrichment.build_openai_client", return_value=_mock_llm_client()):
statuses = []
for _ in range(11):
response = self.client.post("/api/growth_ai_enrichment/run/", payload, format="json")
@@ -591,7 +597,7 @@ class TestAIEnrichmentRunClassification(NonAtomicAPIBaseTest):
payload={"name": "Acme", "contactEmail": "[email protected]"},
)
- with patch("products.growth.backend.api.ai_enrichment.get_llm_client", return_value=_mock_llm_client()):
+ with patch("products.growth.backend.api.ai_enrichment.build_openai_client", return_value=_mock_llm_client()):
response = self.client.post(
"/api/growth_ai_enrichment/run/",
{
@@ -633,7 +639,7 @@ class TestAIEnrichmentRunClassification(NonAtomicAPIBaseTest):
payload={"companyFound": False, "name": "Acme"},
)
- with patch("products.growth.backend.api.ai_enrichment.get_llm_client", return_value=_mock_llm_client()):
+ with patch("products.growth.backend.api.ai_enrichment.build_openai_client", return_value=_mock_llm_client()):
response = self.client.post(
"/api/growth_ai_enrichment/run/",
{
@@ -658,7 +664,7 @@ class TestAIEnrichmentRunClassification(NonAtomicAPIBaseTest):
organization=self.organization, provider="harmonic", payload={"companyFound": False}
)
- with patch("products.growth.backend.api.ai_enrichment.get_llm_client", return_value=_mock_llm_client()):
+ with patch("products.growth.backend.api.ai_enrichment.build_openai_client", return_value=_mock_llm_client()):
response = self.client.post(
"/api/growth_ai_enrichment/run/",
{
@@ -692,7 +698,7 @@ class TestAIEnrichmentRunClassification(NonAtomicAPIBaseTest):
)
client = _mock_llm_client()
- with patch("products.growth.backend.api.ai_enrichment.get_llm_client", return_value=client):
+ with patch("products.growth.backend.api.ai_enrichment.build_openai_client", return_value=client):
response = self.client.post(
"/api/growth_ai_enrichment/run/",
{
@@ -729,7 +735,7 @@ class TestAIEnrichmentRunClassification(NonAtomicAPIBaseTest):
# classify_fetch_for_run runs in a thread-pool worker that only executes (and only then
# calls capture_exception) once something actually iterates streaming_content.
with (
- patch("products.growth.backend.api.ai_enrichment.get_llm_client", return_value=failing_client),
+ patch("products.growth.backend.api.ai_enrichment.build_openai_client", return_value=failing_client),
patch("products.growth.backend.enrichment.lab.capture_exception") as mock_capture,
patch("tenacity.nap.time.sleep"),
):
@@ -765,7 +771,7 @@ class TestAIEnrichmentRunClassification(NonAtomicAPIBaseTest):
client = _mock_llm_client()
with (
- patch("products.growth.backend.api.ai_enrichment.get_llm_client", return_value=client),
+ patch("products.growth.backend.api.ai_enrichment.build_openai_client", return_value=client),
patch("products.growth.backend.enrichment.lab.classify_payload", side_effect=TransientToolError("boom")),
):
response = self.client.post(
diff --git a/products/growth/backend/enrichment/lab.py b/products/growth/backend/enrichment/lab.py
index 02e9d790f287..ba4a3aa8d24e 100644
--- a/products/growth/backend/enrichment/lab.py
+++ b/products/growth/backend/enrichment/lab.py
@@ -18,7 +18,7 @@ import structlog
from openai import OpenAI
from posthog.exceptions_capture import capture_exception
-from posthog.llm.gateway_client import get_llm_client
+from posthog.llm.gateway_client import build_openai_client
from products.growth.backend.enrichment.labels import (
TransientToolError,
@@ -64,7 +64,7 @@ def list_gateway_models() -> list[str]:
if _model_list_cache["models"] is not None and now < _model_list_cache["expires_at"]:
return list(_model_list_cache["models"])
try:
- client = get_llm_client(product="growth")
+ client = build_openai_client(product="growth", ai_product="growth")
models = sorted({model.id for model in client.models.list()})
if not models:
raise ValueError("gateway returned no models")
diff --git a/products/growth/backend/enrichment/labels.py b/products/growth/backend/enrichment/labels.py
index 3be4fc56cb40..fd31266ba8f7 100644
--- a/products/growth/backend/enrichment/labels.py
+++ b/products/growth/backend/enrichment/labels.py
@@ -1,25 +1,32 @@
"""Classifier runner for versioned AI enrichment labels.
Pure functions only: no persistence, no client construction. Callers build the OpenAI
-client via `get_llm_client(product="growth")` and pass it in — this module just turns
-an archived Harmonic payload plus a prompt config into a stamped verdict.
+client via `build_openai_client(product="growth", ai_product="growth", properties=...)` and pass it in —
+this module just turns an archived Harmonic payload plus a prompt config into a stamped verdict.
"""
import re
import json
import math
+import weakref
+import threading
from collections.abc import Callable
from datetime import timedelta
from typing import Any, Literal, TypeIs, cast
from urllib.parse import urlsplit, urlunsplit
+from django.conf import settings
from django.db.models import QuerySet
from django.utils import timezone
+import httpx
+import anthropic
+from anthropic.types import Message
from openai import APIConnectionError, InternalServerError, OpenAI, RateLimitError
from openai.types.chat import ChatCompletion, ChatCompletionMessageParam
from tenacity import retry, retry_if_exception_type, stop_after_attempt, wait_exponential, wait_random
+from posthog.llm.gateway_client import anthropic_gateway_base_url
from posthog.llm.semantic_enrichment import extract_json_object
from posthog.models.organization import Organization, OrganizationMembership
@@ -55,6 +62,8 @@ MAX_INPUT_DEPTH = 6
# not a reservation, so the headroom is free.
MAX_OUTPUT_TOKENS = 4000
+AI_ENRICHMENT_FEATURE = "ai_enrichment"
+
# Bounds on the prompt config's own authored content (prompt_text, output_fields) - not the
# archived payload data above. Defined here, not in ai_enrichment_serializers.py, so any other
# consumer of EnrichmentPromptConfig content (e.g. the admin form) can share the same numbers
@@ -459,6 +468,28 @@ def _run_tool_call(
return run_tool(call.function.name, arguments, pacer=pacer)
+_RETRYABLE_ERRORS = (
+ APIConnectionError,
+ RateLimitError,
+ InternalServerError,
+ anthropic.APIConnectionError,
+ anthropic.RateLimitError,
+ anthropic.InternalServerError,
+)
+
+_ANTHROPIC_TOOLS = [
+ {
+ "name": tool["function"]["name"],
+ "description": tool["function"]["description"],
+ "input_schema": tool["function"]["parameters"],
+ }
+ for tool in TOOLS
+]
+
+_anthropic_clients: "weakref.WeakKeyDictionary[OpenAI, anthropic.Anthropic]" = weakref.WeakKeyDictionary()
+_anthropic_clients_lock = threading.Lock()
+
+
@retry(
# Allowlist, not a blacklist: only failures that can plausibly succeed on a retry earn one.
# APIConnectionError covers its APITimeoutError subclass too. Everything else — most
@@ -467,16 +498,130 @@ def _run_tool_call(
# three attempts from stacking with the SDK's own retries on a hard failure like a bad
# gateway credential; see max_retries=0 at every client construction site for the other half
# of that fix.
- retry=retry_if_exception_type((APIConnectionError, RateLimitError, InternalServerError)),
+ retry=retry_if_exception_type(_RETRYABLE_ERRORS),
stop=stop_after_attempt(3),
# Jitter so a worker pool that all hit a 429 together doesn't retry in lockstep.
wait=wait_exponential(multiplier=1, min=2, max=30) + wait_random(0, 2),
reraise=True,
)
def _complete(client: OpenAI, request: dict[str, Any]) -> ChatCompletion:
+ if _routes_through_anthropic_messages(client, request["model"]):
+ return _complete_with_anthropic_messages(client, request)
return client.chat.completions.create(**request)
+def _routes_through_anthropic_messages(client: OpenAI, model: str) -> bool:
+ if not model.startswith("claude-"):
+ return False
+ if not settings.AI_GATEWAY_URL:
+ return False
+ return str(client.base_url).rstrip("/") == settings.AI_GATEWAY_URL.rstrip("/")
+
+
+def _anthropic_client_for(client: OpenAI) -> anthropic.Anthropic:
+ with _anthropic_clients_lock:
+ anthropic_client = _anthropic_clients.get(client)
+ if anthropic_client is None:
+ anthropic_client = anthropic.Anthropic(
+ api_key=client.api_key,
+ base_url=anthropic_gateway_base_url(str(client.base_url)),
+ default_headers={
+ name: value
+ for name, value in client.default_headers.items()
+ if name.lower().startswith("x-posthog-") and isinstance(value, str)
+ },
+ max_retries=client.max_retries,
+ http_client=httpx.Client(trust_env=False),
+ )
+ _anthropic_clients[client] = anthropic_client
+ return anthropic_client
+
+
+def _complete_with_anthropic_messages(client: OpenAI, request: dict[str, Any]) -> ChatCompletion:
+ message = _anthropic_client_for(client).messages.create(**_anthropic_request(request))
+ return _chat_completion_from_message(message)
+
+
+def _anthropic_request(request: dict[str, Any]) -> dict[str, Any]:
+ system, *history = request["messages"]
+ assert system["role"] == "system"
+ assert history
+ messages: list[dict[str, Any]] = []
+ for m in history:
+ if m["role"] == "tool":
+ result = {"type": "tool_result", "tool_use_id": m["tool_call_id"], "content": m["content"]}
+ if messages[-1]["role"] == "user":
+ messages[-1]["content"].append(result)
+ else:
+ messages.append({"role": "user", "content": [result]})
+ continue
+ if m["role"] == "assistant" and m.get("tool_calls"):
+ blocks: list[dict[str, Any]] = [{"type": "text", "text": m["content"]}] if m.get("content") else []
+ blocks.extend(
+ {
+ "type": "tool_use",
+ "id": call["id"],
+ "name": call["function"]["name"],
+ "input": json.loads(call["function"]["arguments"]),
+ }
+ for call in m["tool_calls"]
+ )
+ messages.append({"role": "assistant", "content": blocks})
+ continue
+ messages.append({"role": m["role"], "content": m["content"]})
+ anthropic_request: dict[str, Any] = {
+ "model": request["model"],
+ "system": system["content"],
+ "messages": messages,
+ "max_tokens": request["max_completion_tokens"],
+ "timeout": request["timeout"],
+ }
+ if "tools" in request:
+ assert request["tools"] is TOOLS
+ anthropic_request["tools"] = _ANTHROPIC_TOOLS
+ elif any(m["role"] == "tool" for m in history):
+ anthropic_request["tools"] = _ANTHROPIC_TOOLS
+ anthropic_request["tool_choice"] = {"type": "none"}
+ return anthropic_request
+
+
+_ANTHROPIC_FINISH_REASONS = {
+ "end_turn": "stop",
+ "stop_sequence": "stop",
+ "max_tokens": "length",
+ "tool_use": "tool_calls",
+}
+
+
+def _chat_completion_from_message(message: Message) -> ChatCompletion:
+ text = "".join(block.text for block in message.content if block.type == "text")
+ tool_calls = [
+ {"id": block.id, "type": "function", "function": {"name": block.name, "arguments": json.dumps(block.input)}}
+ for block in message.content
+ if block.type == "tool_use"
+ ]
+ return ChatCompletion.model_validate(
+ {
+ "id": message.id,
+ "object": "chat.completion",
+ "created": int(timezone.now().timestamp()),
+ "model": message.model,
+ "choices": [
+ {
+ "index": 0,
+ "finish_reason": _ANTHROPIC_FINISH_REASONS.get(message.stop_reason or "", "stop"),
+ "message": {"role": "assistant", "content": text or None, "tool_calls": tool_calls or None},
+ }
+ ],
+ "usage": {
+ "prompt_tokens": message.usage.input_tokens,
+ "completion_tokens": message.usage.output_tokens,
+ "total_tokens": message.usage.input_tokens + message.usage.output_tokens,
+ },
+ }
+ )
+
+
def _call_and_parse(
config: EnrichmentPromptConfig,
messages: list[dict[str, Any]],
diff --git a/products/growth/backend/management/commands/enrichment_label_batch.py b/products/growth/backend/management/commands/enrichment_label_batch.py
index 723c87f90fb0..f7574efcd245 100644
--- a/products/growth/backend/management/commands/enrichment_label_batch.py
+++ b/products/growth/backend/management/commands/enrichment_label_batch.py
@@ -24,7 +24,7 @@ from django.db.models import Q
import structlog
from posthog.exceptions_capture import capture_exception
-from posthog.llm.gateway_client import get_llm_client
+from posthog.llm.gateway_client import build_openai_client
from posthog.ph_client import ph_scoped_capture
from posthog.utils import get_instance_region
@@ -32,6 +32,7 @@ from products.growth.backend.enrichment import gates
from products.growth.backend.enrichment.fit_recomputation import apply_enrichment_result, label_needs_application
from products.growth.backend.enrichment.icp_lists import load_active_lists
from products.growth.backend.enrichment.labels import (
+ AI_ENRICHMENT_FEATURE,
PromptConfigError,
TransientToolError,
ai_processing_approved,
@@ -198,7 +199,11 @@ class Command(BaseCommand):
# tenacity in labels.py already owns retries (stop_after_attempt(3)); the SDK's own
# internal retries underneath would multiply that budget nine-fold per fetch and actively
# worsen a 429 the tenacity layer is already backing off from.
- client = get_llm_client(product="growth").with_options(max_retries=0)
+ client = build_openai_client(
+ product="growth",
+ ai_product="growth",
+ properties={"ai_feature": AI_ENRICHMENT_FEATURE, "surface": "batch"},
+ ).with_options(max_retries=0)
pacer = FirecrawlPacer()
counts: dict[str, int] = {
diff --git a/products/growth/backend/management/commands/enrichment_label_dry_run.py b/products/growth/backend/management/commands/enrichment_label_dry_run.py
index 3a0e0693d5e5..8bc09e0e9d53 100644
--- a/products/growth/backend/management/commands/enrichment_label_dry_run.py
+++ b/products/growth/backend/management/commands/enrichment_label_dry_run.py
@@ -9,9 +9,10 @@ from typing import Any
from django.core.management.base import BaseCommand, CommandError, CommandParser
-from posthog.llm.gateway_client import get_llm_client
+from posthog.llm.gateway_client import build_openai_client
from products.growth.backend.enrichment.labels import (
+ AI_ENRICHMENT_FEATURE,
UNKNOWN,
PromptConfigError,
TransientToolError,
@@ -94,7 +95,11 @@ class Command(BaseCommand):
# tenacity in labels.py already owns retries; the SDK's own internal retries underneath
# would multiply that budget nine-fold per row.
- client = get_llm_client(product="growth").with_options(max_retries=0)
+ client = build_openai_client(
+ product="growth",
+ ai_product="growth",
+ properties={"ai_feature": AI_ENRICHMENT_FEATURE, "surface": "dry_run"},
+ ).with_options(max_retries=0)
# A custom output schema's pass/fail key differs from `label` - see
# verdict_field_key's docstring.
verdict_key = verdict_field_key(config)
diff --git a/products/growth/backend/tests/test_ai_pilled.py b/products/growth/backend/tests/test_ai_pilled.py
index 62510613d542..3988a274be71 100644
--- a/products/growth/backend/tests/test_ai_pilled.py
+++ b/products/growth/backend/tests/test_ai_pilled.py
@@ -294,7 +294,7 @@ class TestEnrichmentScoreApplication(BaseTest):
return_value=OrganizationBridgeInputs(),
),
patch(
- "products.growth.backend.management.commands.enrichment_label_batch.get_llm_client",
+ "products.growth.backend.management.commands.enrichment_label_batch.build_openai_client",
return_value=gateway,
),
):
@@ -357,7 +357,7 @@ class TestEnrichmentScoreApplication(BaseTest):
return_value=OrganizationBridgeInputs(),
),
patch(
- "products.growth.backend.management.commands.enrichment_label_batch.get_llm_client",
+ "products.growth.backend.management.commands.enrichment_label_batch.build_openai_client",
return_value=gateway,
),
patch(
@@ -499,7 +499,7 @@ class TestEnrichmentScoreApplication(BaseTest):
return_value=OrganizationBridgeInputs(),
),
patch(
- "products.growth.backend.management.commands.enrichment_label_batch.get_llm_client",
+ "products.growth.backend.management.commands.enrichment_label_batch.build_openai_client",
return_value=gateway,
),
):
@@ -547,7 +547,7 @@ class TestEnrichmentScoreApplication(BaseTest):
return_value=OrganizationBridgeInputs(),
),
patch(
- "products.growth.backend.management.commands.enrichment_label_batch.get_llm_client",
+ "products.growth.backend.management.commands.enrichment_label_batch.build_openai_client",
return_value=gateway,
),
patch("products.growth.backend.management.commands.enrichment_label_batch._ID_BATCH_SIZE", 1),
@@ -599,7 +599,7 @@ class TestEnrichmentScoreApplication(BaseTest):
return_value=OrganizationBridgeInputs(),
),
patch(
- "products.growth.backend.management.commands.enrichment_label_batch.get_llm_client",
+ "products.growth.backend.management.commands.enrichment_label_batch.build_openai_client",
return_value=gateway,
),
patch("products.growth.backend.management.commands.enrichment_label_batch._ID_BATCH_SIZE", 1),
diff --git a/products/growth/backend/tests/test_enrichment_label_commands.py b/products/growth/backend/tests/test_enrichment_label_commands.py
index 4e26b3aba6b9..c96197dc50bd 100644
--- a/products/growth/backend/tests/test_enrichment_label_commands.py
+++ b/products/growth/backend/tests/test_enrichment_label_commands.py
@@ -56,7 +56,7 @@ def _bad_response() -> MagicMock:
def _mock_llm_client() -> MagicMock:
client = MagicMock()
- # Command call sites chain .with_options(max_retries=0) onto get_llm_client(...); without
+ # Command call sites chain .with_options(max_retries=0) onto build_openai_client(...); without
# this self-reference, that call returns an unconfigured child mock and every call made
# through it is invisible to assertions on `client`.
client.with_options.return_value = client
@@ -96,14 +96,14 @@ class _BatchCommandTestCase(BaseTest):
class TestGatewayRetryBudget(_BatchCommandTestCase):
def test_batch_disables_sdk_internal_retries(self):
- # get_llm_client builds an OpenAI() client without max_retries, so the SDK's own default
+ # build_openai_client builds an OpenAI() client without max_retries, so the SDK's own default
# of 2 stacks under tenacity's stop_after_attempt(3): up to 9 HTTP requests per fetch
# instead of 3. The command must opt out of the SDK's layer so tenacity is the only one.
self._config()
self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
client.with_options.assert_called_once_with(max_retries=0)
@@ -113,7 +113,7 @@ class TestGatewayRetryBudget(_BatchCommandTestCase):
self._fetch()
client = _mock_llm_client()
- with patch(f"{_DRY_RUN_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_DRY_RUN_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_dry_run", label="test_label", sample=1)
client.with_options.assert_called_once_with(max_retries=0)
@@ -128,7 +128,7 @@ class TestCircuitBreaker(_BatchCommandTestCase):
client.chat.completions.create.return_value = _bad_response()
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.capture_exception"),
):
with self.assertRaises(CommandError) as ctx:
@@ -152,7 +152,7 @@ class TestCircuitBreaker(_BatchCommandTestCase):
]
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.capture_exception"),
):
# max_failures=2 would trip on two failures in a row; these two never land back to
@@ -180,7 +180,7 @@ class TestKeysetPagination(_BatchCommandTestCase):
options["limit"] = limit
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
# Forces the 5 orgs across 3 keyset pages instead of 1, so a cursor off-by-one would
# duplicate or skip a row instead of silently passing on a single-page test.
patch(f"{_BATCH_COMMAND_MODULE}._ID_BATCH_SIZE", 2),
@@ -216,7 +216,7 @@ class TestKeysetPagination(_BatchCommandTestCase):
out = StringIO()
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
# Forces the 3 already-done orgs and the 1 fresh org across separate keyset pages.
patch(f"{_BATCH_COMMAND_MODULE}._ID_BATCH_SIZE", 2),
):
@@ -243,7 +243,7 @@ class TestKeysetPagination(_BatchCommandTestCase):
client = _mock_llm_client()
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}._ID_BATCH_SIZE", 2),
):
call_command("enrichment_label_batch", label="test_label", workers=1, **options)
@@ -282,7 +282,7 @@ class TestAdvisoryLock(_BatchCommandTestCase):
assert holder_ready.wait(timeout=5)
assert holder_state["acquired"] is True
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
with self.assertRaises(CommandError) as ctx:
call_command("enrichment_label_batch", label="test_label", workers=1)
assert "already holds the lock" in str(ctx.exception)
@@ -304,7 +304,7 @@ class TestUnknownAccounting(_BatchCommandTestCase):
client = _mock_llm_client()
out = StringIO()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1, stdout=out)
assert "unknown 1" in out.getvalue()
@@ -336,7 +336,7 @@ class TestExitCodeAndSummary(_BatchCommandTestCase):
options["min_success_rate"] = min_success_rate
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.capture_exception"),
):
if expect_raise:
@@ -359,7 +359,7 @@ class TestExitCodeAndSummary(_BatchCommandTestCase):
out = StringIO()
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.capture_exception"),
):
with self.assertRaises(CommandError):
@@ -379,7 +379,7 @@ class TestExitCodeAndSummary(_BatchCommandTestCase):
]
out = StringIO()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1, stdout=out)
assert "prompt_tokens 350" in out.getvalue()
@@ -401,7 +401,7 @@ class TestExpectedVersionGuard(_BatchCommandTestCase):
self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
with self.assertRaises(CommandError):
call_command("enrichment_label_batch", label="test_label", workers=1, expected_version="v1")
@@ -413,7 +413,7 @@ class TestExpectedVersionGuard(_BatchCommandTestCase):
self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1, expected_version="v2")
assert EnrichmentLabelResult.objects.count() == 1
@@ -423,7 +423,7 @@ class TestExpectedVersionGuard(_BatchCommandTestCase):
self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
assert EnrichmentLabelResult.objects.count() == 1
@@ -444,7 +444,7 @@ class TestFirecrawlKeyGuard(_BatchCommandTestCase):
with (
override_settings(CLOUD_DEPLOYMENT=cloud_deployment, FIRECRAWL_API_KEY=firecrawl_key),
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
):
if aborts:
with self.assertRaisesMessage(CommandError, "FIRECRAWL_API_KEY"):
@@ -477,7 +477,7 @@ class TestWorkerConnectionErrors(NonAtomicBaseTest):
client = _mock_llm_client()
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.close_old_connections", side_effect=RuntimeError("connection reset")),
patch(f"{_BATCH_COMMAND_MODULE}.capture_exception") as capture_mock,
):
@@ -511,7 +511,7 @@ class TestDryRunFixes(BaseTest):
client = _mock_llm_client()
out = StringIO()
- with patch(f"{_DRY_RUN_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_DRY_RUN_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_dry_run", label="test_label", sample=2, stdout=out)
printed = out.getvalue()
@@ -525,7 +525,7 @@ class TestDryRunFixes(BaseTest):
)
client = _mock_llm_client()
- with patch(f"{_DRY_RUN_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_DRY_RUN_COMMAND_MODULE}.build_openai_client", return_value=client):
with self.assertRaises(CommandError):
call_command("enrichment_label_dry_run", label="test_label", sample=1)
@@ -551,7 +551,7 @@ class TestDryRunToolDeferral(BaseTest):
out = StringIO()
with (
- patch(f"{_DRY_RUN_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_DRY_RUN_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_DRY_RUN_COMMAND_MODULE}.classify_payload", side_effect=TransientToolError("boom")),
):
call_command("enrichment_label_dry_run", label="test_label", sample=1, stdout=out)
@@ -580,7 +580,7 @@ class TestToolCallAccounting(_BatchCommandTestCase):
}
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.classify_payload", side_effect=_classify),
):
call_command("enrichment_label_batch", label="test_label", workers=1, stdout=out)
@@ -596,7 +596,7 @@ class TestToolDeferral(_BatchCommandTestCase):
out = StringIO()
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.classify_payload", side_effect=TransientToolError("boom")),
self.assertRaises(CommandError),
):
@@ -625,7 +625,7 @@ class TestToolDeferral(_BatchCommandTestCase):
out = StringIO()
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_TOOLS_MODULE}.get_outbound_rate_limiter", return_value=limiter),
patch(f"{_TOOLS_MODULE}.time.sleep"),
patch(
@@ -646,7 +646,7 @@ class TestToolDeferral(_BatchCommandTestCase):
out = StringIO()
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.classify_payload", side_effect=TransientToolError("boom")) as classify_mock,
self.assertRaises(CommandError) as ctx,
):
@@ -673,7 +673,7 @@ class TestToolDeferral(_BatchCommandTestCase):
return {"is_ai": True, "confidence": 0.9, "reasoning": "x", "inputs": {"signup_domain": None, "fields": {}}}
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.classify_payload", side_effect=_classify),
self.assertRaises(CommandError) as ctx,
):
@@ -695,7 +695,7 @@ class TestBatchRunReportEvent(_BatchCommandTestCase):
scoped_capture.__enter__.return_value = capture
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.get_instance_region", return_value="EU"),
patch(f"{_BATCH_COMMAND_MODULE}.ph_scoped_capture", return_value=scoped_capture) as scoped_capture_factory,
):
@@ -724,7 +724,7 @@ class TestBatchRunReportEvent(_BatchCommandTestCase):
out = StringIO()
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.get_instance_region", return_value="US"),
patch(f"{_BATCH_COMMAND_MODULE}.ph_scoped_capture", side_effect=RuntimeError("boom")),
patch(f"{_BATCH_COMMAND_MODULE}.capture_exception") as capture_mock,
@@ -741,7 +741,7 @@ class TestBatchRunReportEvent(_BatchCommandTestCase):
client = _mock_llm_client()
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.get_instance_region", return_value=None),
patch(f"{_BATCH_COMMAND_MODULE}.ph_scoped_capture") as scoped_capture_factory,
):
@@ -758,7 +758,7 @@ class TestBatchRunReportEvent(_BatchCommandTestCase):
scoped_capture.__enter__.return_value = MagicMock()
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.capture_exception"),
patch(f"{_BATCH_COMMAND_MODULE}.get_instance_region", return_value="US"),
patch(f"{_BATCH_COMMAND_MODULE}.ph_scoped_capture", return_value=scoped_capture) as scoped_capture_factory,
@@ -777,7 +777,7 @@ class TestAiProcessingConsent(_BatchCommandTestCase):
self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
client.chat.completions.create.assert_not_called()
@@ -792,11 +792,11 @@ class TestAiProcessingConsent(_BatchCommandTestCase):
self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
self.organization.is_ai_data_processing_approved = True
self.organization.save(update_fields=["is_ai_data_processing_approved"])
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
assert EnrichmentLabelResult.objects.count() == 1
@@ -817,7 +817,7 @@ class TestAiProcessingConsent(_BatchCommandTestCase):
client.chat.completions.create.side_effect = _revoke_after_first
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
assert client.chat.completions.create.call_count == 1
@@ -830,7 +830,7 @@ class TestAiProcessingConsent(_BatchCommandTestCase):
self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
client.chat.completions.create.assert_not_called()
@@ -843,7 +843,7 @@ class TestAiProcessingConsent(_BatchCommandTestCase):
client = _mock_llm_client()
out = StringIO()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1, stdout=out)
assert "skipped_no_ai_consent 1" in out.getvalue()
@@ -866,7 +866,7 @@ class TestAiProcessingConsent(_BatchCommandTestCase):
client = _mock_llm_client()
out = StringIO()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1, limit=1, stdout=out)
assert EnrichmentLabelResult.objects.filter(organization=approved_org).exists()
@@ -882,7 +882,7 @@ class TestAiProcessingConsent(_BatchCommandTestCase):
client = _mock_llm_client()
out = StringIO()
- with patch(f"{_DRY_RUN_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_DRY_RUN_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_dry_run", label="test_label", sample=1, stdout=out)
output = out.getvalue()
diff --git a/products/growth/backend/tests/test_enrichment_labels.py b/products/growth/backend/tests/test_enrichment_labels.py
index fd46ddd4d565..e33d001f84c8 100644
--- a/products/growth/backend/tests/test_enrichment_labels.py
+++ b/products/growth/backend/tests/test_enrichment_labels.py
@@ -9,9 +9,10 @@ from unittest.mock import MagicMock, patch
from django.core.management import call_command
from django.core.management.base import CommandError
-from django.test import SimpleTestCase
+from django.test import SimpleTestCase, override_settings
import openai
+from anthropic.types import Message, TextBlock, ToolUseBlock, Usage
from openai import OpenAI
from parameterized import parameterized
@@ -43,7 +44,7 @@ def _mock_llm_client(
verdict: bool = True, confidence: float = 0.9, reasoning: str = "builds ai software", verdict_key: str = "is_ai"
) -> MagicMock:
client = MagicMock()
- # The command call sites chain .with_options(max_retries=0) onto get_llm_client(...); without
+ # The command call sites chain .with_options(max_retries=0) onto build_openai_client(...); without
# this, that call returns a fresh, unconfigured child mock and every assertion below silently
# checks the wrong object.
client.with_options.return_value = client
@@ -403,6 +404,104 @@ class TestClassifyPayloadToolLoop(SimpleTestCase):
assert len(client.calls) == 3
+ @override_settings(AI_GATEWAY_URL="https://gateway.test/v1")
+ def test_claude_on_the_ai_gateway_runs_the_tool_loop_through_anthropic_messages(self):
+ config = self._config()
+ config.model = "claude-haiku-4-5"
+ client = OpenAI(
+ api_key="phs_test", base_url="https://gateway.test/v1", default_headers={"X-PostHog-Product": "growth"}
+ )
+ usage = Usage(input_tokens=10, output_tokens=5)
+ replies = [
+ Message(
+ id="msg_1",
+ type="message",
+ role="assistant",
+ model="claude-haiku-4-5",
+ stop_reason="tool_use",
+ content=[ToolUseBlock(type="tool_use", id="toolu_1", name="web_search", input={"query": "Acme AI"})],
+ usage=usage,
+ ),
+ Message(
+ id="msg_2",
+ type="message",
+ role="assistant",
+ model="claude-haiku-4-5",
+ stop_reason="end_turn",
+ content=[TextBlock(type="text", text=json.dumps({"is_ai": True, "evidence_url": ""}))],
+ usage=usage,
+ ),
+ ]
+ found = FirecrawlSearch(
+ query="Acme AI", results=(FirecrawlSearchResult(url="https://techcrunch.com/acme", title="funding"),)
+ )
+
+ with (
+ patch(f"{_LABELS_MODULE}.anthropic.Anthropic") as anthropic_client,
+ patch(f"{_TOOLS_MODULE}.search", return_value=found),
+ ):
+ create = anthropic_client.return_value.messages.create
+ create.side_effect = replies
+ result = classify_payload(config, {"name": "Acme"}, "example.com", client)
+
+ assert result["is_ai"] is True
+ assert result["meta"]["prompt_tokens"] == 20
+ assert anthropic_client.call_args.kwargs["base_url"] == "https://gateway.test"
+ assert anthropic_client.call_args.kwargs["default_headers"] == {"X-PostHog-Product": "growth"}
+ first_turn, second_turn = (call.kwargs for call in create.call_args_list)
+ assert first_turn["system"].startswith("judge it.")
+ assert [tool["name"] for tool in first_turn["tools"]] == ["web_search", "fetch_page"]
+ assert second_turn["messages"][1]["content"][0]["type"] == "tool_use"
+ tool_result = second_turn["messages"][2]["content"][0]
+ assert tool_result["tool_use_id"] == "toolu_1"
+ assert "techcrunch.com/acme" in tool_result["content"]
+
+ @override_settings(AI_GATEWAY_URL="https://gateway.test/v1")
+ def test_claude_final_turn_after_the_tool_budget_keeps_tools_but_disables_them(self):
+ config = self._config()
+ config.model = "claude-haiku-4-5"
+ client = OpenAI(api_key="phs_test", base_url="https://gateway.test/v1")
+ usage = Usage(input_tokens=10, output_tokens=5)
+ tool_uses = [
+ ToolUseBlock(type="tool_use", id=f"toolu_{i}", name="web_search", input={"query": f"q{i}"})
+ for i in range(5)
+ ]
+ replies = [
+ Message(
+ id="msg_1",
+ type="message",
+ role="assistant",
+ model="claude-haiku-4-5",
+ stop_reason="tool_use",
+ content=tool_uses,
+ usage=usage,
+ ),
+ Message(
+ id="msg_2",
+ type="message",
+ role="assistant",
+ model="claude-haiku-4-5",
+ stop_reason="end_turn",
+ content=[TextBlock(type="text", text=json.dumps({"is_ai": True, "evidence_url": ""}))],
+ usage=usage,
+ ),
+ ]
+ found = FirecrawlSearch(query="q", results=(FirecrawlSearchResult(url="https://x.example"),))
+
+ with (
+ patch(f"{_LABELS_MODULE}.anthropic.Anthropic") as anthropic_client,
+ patch(f"{_TOOLS_MODULE}.search", return_value=found),
+ ):
+ create = anthropic_client.return_value.messages.create
+ create.side_effect = replies
+ result = classify_payload(config, {"name": "Acme"}, "example.com", client)
+
+ assert result["is_ai"] is True
+ final_turn = create.call_args_list[1].kwargs
+ assert [tool["name"] for tool in final_turn["tools"]] == ["web_search", "fetch_page"]
+ assert final_turn["tool_choice"] == {"type": "none"}
+ assert len(final_turn["messages"][-1]["content"]) == 5
+
class TestClassifyPayloadToolEvidenceUrl(SimpleTestCase):
def _config(self) -> EnrichmentPromptConfig:
@@ -776,7 +875,7 @@ class TestEnrichmentLabelBatch(BaseTest):
fetch = self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
result = EnrichmentLabelResult.objects.get(organization=self.organization, label_name="test_label")
@@ -799,7 +898,7 @@ class TestEnrichmentLabelBatch(BaseTest):
response.usage.completion_tokens = 40
client.chat.completions.create.return_value = response
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
output = EnrichmentLabelResult.objects.get(label_name="test_label").output
@@ -819,7 +918,7 @@ class TestEnrichmentLabelBatch(BaseTest):
self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
result = EnrichmentLabelResult.objects.get(label_name="test_label")
@@ -831,7 +930,7 @@ class TestEnrichmentLabelBatch(BaseTest):
self._fetch(payload={})
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
result = EnrichmentLabelResult.objects.get(label_name="test_label")
@@ -845,7 +944,7 @@ class TestEnrichmentLabelBatch(BaseTest):
self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
call_command("enrichment_label_batch", label="test_label", workers=1)
@@ -857,7 +956,7 @@ class TestEnrichmentLabelBatch(BaseTest):
first_fetch = self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
second_fetch = self._fetch(payload={"name": "Acme v2"})
@@ -865,7 +964,7 @@ class TestEnrichmentLabelBatch(BaseTest):
fetched_at=first_fetch.fetched_at + dt.timedelta(minutes=5)
)
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
rows = EnrichmentLabelResult.objects.filter(
@@ -890,7 +989,7 @@ class TestEnrichmentLabelBatch(BaseTest):
return original_classify(*args, **kwargs)
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.classify_payload", side_effect=rename_then_classify),
):
call_command("enrichment_label_batch", label="test_label", workers=1)
@@ -916,7 +1015,7 @@ class TestEnrichmentLabelBatch(BaseTest):
client.chat.completions.create.return_value = response
with (
- patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client),
+ patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client),
patch(f"{_BATCH_COMMAND_MODULE}.capture_exception") as capture_mock,
patch("tenacity.nap.time.sleep"),
):
@@ -950,7 +1049,7 @@ class TestEnrichmentLabelBatch(BaseTest):
client = MagicMock()
out = StringIO()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1, stdout=out)
assert "unknown 1" in out.getvalue()
@@ -963,14 +1062,14 @@ class TestEnrichmentLabelBatch(BaseTest):
self._fetch()
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
v1.is_active = False
v1.save()
self._config(version="ai-pilled-clay-v2", prompt_text="a different prompt entirely. Email: {email}")
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=1)
assert EnrichmentLabelResult.objects.filter(prompt_version="ai-pilled-clay-v1").count() == 1
@@ -1016,7 +1115,7 @@ class TestEnrichmentLabelDryRun(BaseTest):
)
client = _mock_llm_client()
- with patch(f"{_DRY_RUN_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_DRY_RUN_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_dry_run", label="test_label")
assert EnrichmentLabelResult.objects.count() == 0
@@ -1068,7 +1167,7 @@ class TestEnrichmentLabelDryRun(BaseTest):
client.chat.completions.create.return_value = response
out = StringIO()
- with patch(f"{_DRY_RUN_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_DRY_RUN_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command(
"enrichment_label_dry_run",
label="ai_native_teams",
@@ -1113,13 +1212,13 @@ class TestEnrichmentLabelBatchConcurrency(NonAtomicBaseTest):
)
client = _mock_llm_client()
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=2)
assert client.chat.completions.create.call_count == 4
assert EnrichmentLabelResult.objects.count() == 4
- with patch(f"{_BATCH_COMMAND_MODULE}.get_llm_client", return_value=client):
+ with patch(f"{_BATCH_COMMAND_MODULE}.build_openai_client", return_value=client):
call_command("enrichment_label_batch", label="test_label", workers=2)
assert client.chat.completions.create.call_count == 4