fix(router): prefer web search deployments at the strategy pick and fall back when none has capacity

This commit is contained in:
mateo-berri 2026-10-29 13:13:54 -07:00
parent 5b2ccc8e58
commit a179118b3e
3 changed files with 252 additions and 40 deletions

View file

@ -235,6 +235,7 @@ from litellm.router_utils.clientside_credential_handler import (
)
from litellm.router_utils.common_utils import (
MID_STREAM_CONTINUATION_KWARG,
apick_preferring_web_search,
as_retry_skipped_deployment_ids,
deployments_the_caller_may_use,
filter_team_based_models,
@ -245,6 +246,7 @@ from litellm.router_utils.common_utils import (
format_no_fallback_group_message,
get_request_team_id,
is_proxy_admin_request,
pick_preferring_web_search,
prefer_web_search_deployments,
provider_for_generic_call,
request_user_api_key_auth,
@ -15246,13 +15248,10 @@ class Router:
## ORDER FILTERING ## -> if user set 'order' in deployments, return deployments with lowest order (e.g. order=1 > order=2)
_target_order: Final = (request_kwargs or {}).pop("_target_order", None)
batch_candidates: Final = prefer_web_search_deployments(
self._without_declared_text_only_deployments(
cast(list[dict[str, object]], healthy_deployments), # cast-ok: router deployments are plain dicts
model,
messages,
cast(Mapping[str, object] | None, request_kwargs), # cast-ok: legacy signature uses a bare dict
),
batch_candidates: Final = self._without_declared_text_only_deployments(
cast(list[dict[str, object]], healthy_deployments), # cast-ok: router deployments are plain dicts
model,
messages,
cast(Mapping[str, object] | None, request_kwargs), # cast-ok: legacy signature uses a bare dict
)
embedding_batch_admitting: Final = deployments_admitting_embedding_batch(
@ -15535,26 +15534,31 @@ class Router:
return healthy_deployments[0]
start_time: Final = time.time()
routing_kwargs: Final = cast(Mapping[str, object], request_kwargs) # cast-ok: str-keyed request kwargs
if strategy == "simple-shuffle":
shuffled: Final = simple_shuffle(
resolve_model_alias=self.get_model_from_alias,
healthy_deployments=healthy_deployments,
healthy_deployments=prefer_web_search_deployments(healthy_deployments, routing_kwargs),
model=model,
request_kwargs=request_kwargs,
request_kwargs=routing_kwargs,
)
phase_event(_DEPLOYMENT_SELECTED_EVENT, pick_attributes)
return shuffled
with PrefetchedUsage.scoped(
routing_read_batch.prefetched_usage if routing_read_batch is not None else None
):
deployment: Final = await self._select_deployment_async(
strategy=strategy,
selector=strategy_selector,
model=model,
healthy_deployments=healthy_deployments,
messages=messages,
input=input,
request_kwargs=request_kwargs,
deployment: Final = await apick_preferring_web_search(
lambda candidates: self._select_deployment_async(
strategy=strategy,
selector=strategy_selector,
model=model,
healthy_deployments=candidates,
messages=messages,
input=input,
request_kwargs=request_kwargs,
),
cast(list[dict[str, object]], healthy_deployments), # cast-ok: router deployments are plain dicts
routing_kwargs,
)
if deployment is None:
exception: Final = await async_raise_no_deployment_exception(
@ -16481,13 +16485,10 @@ class Router:
## ORDER FILTERING ## -> if user set 'order' in deployments, return deployments with lowest order (e.g. order=1 > order=2)
_target_order: Final = (request_kwargs or {}).pop("_target_order", None)
batch_candidates: Final = prefer_web_search_deployments(
self._without_declared_text_only_deployments(
cast(list[dict[str, object]], healthy_deployments), # cast-ok: router deployments are plain dicts
model,
messages,
cast(Mapping[str, object] | None, request_kwargs), # cast-ok: legacy signature uses a bare dict
),
batch_candidates: Final = self._without_declared_text_only_deployments(
cast(list[dict[str, object]], healthy_deployments), # cast-ok: router deployments are plain dicts
model,
messages,
cast(Mapping[str, object] | None, request_kwargs), # cast-ok: legacy signature uses a bare dict
)
embedding_batch_admitting: Final = deployments_admitting_embedding_batch(
@ -16575,20 +16576,24 @@ class Router:
############## Check 'weight' param set for weighted pick #################
shuffled: Final = simple_shuffle(
resolve_model_alias=self.get_model_from_alias,
healthy_deployments=healthy_deployments,
healthy_deployments=prefer_web_search_deployments(healthy_deployments, routing_kwargs),
model=model,
request_kwargs=request_kwargs,
request_kwargs=routing_kwargs,
)
phase_event(_DEPLOYMENT_SELECTED_EVENT, pick_attributes)
return shuffled
deployment: Final = self._select_deployment_sync(
strategy=strategy,
selector=strategy_selector,
model=model,
healthy_deployments=healthy_deployments,
messages=messages,
input=input,
request_kwargs=request_kwargs,
deployment: Final = pick_preferring_web_search(
lambda candidates: self._select_deployment_sync(
strategy=strategy,
selector=strategy_selector,
model=model,
healthy_deployments=candidates,
messages=messages,
input=input,
request_kwargs=request_kwargs,
),
cast(list[dict[str, object]], healthy_deployments), # cast-ok: router deployments are plain dicts
routing_kwargs,
)
if deployment is None:

View file

@ -1,6 +1,6 @@
import hashlib
import json
from collections.abc import Mapping, Sequence
from collections.abc import Awaitable, Callable, Mapping, Sequence
from types import MappingProxyType
from typing import TYPE_CHECKING, Final, TypeVar
@ -33,6 +33,8 @@ from litellm.types.utils import (
)
_V = TypeVar("_V")
_Picked: Final = TypeVar("_Picked")
_DeploymentT: Final = TypeVar("_DeploymentT", bound=dict[str, object])
MID_STREAM_CONTINUATION_KWARG: Final = "_mid_stream_continuation"
ROUTER_ONLY_CALL_KWARGS: Final = frozenset({"silent_model", "include_fallback_errors", MID_STREAM_CONTINUATION_KWARG})
@ -53,6 +55,9 @@ _DEPLOYMENT_ROWS_ADAPTER: Final[TypeAdapter[tuple[Mapping[str, object], ...]]] =
tuple[Mapping[str, object], ...]
)
_DEPLOYMENT_IDS_ADAPTER: Final = TypeAdapter(tuple[str, ...])
_RETRY_BREADCRUMBS_ADAPTER: Final[TypeAdapter[tuple[Mapping[str, object], ...]]] = TypeAdapter(
tuple[Mapping[str, object], ...]
)
_FALLBACK_ENTRIES_ADAPTER: Final[TypeAdapter[tuple[Mapping[str, object], ...]]] = TypeAdapter(
tuple[Mapping[str, object], ...]
)
@ -493,10 +498,31 @@ def _deployment_id(deployment: Mapping[str, object]) -> object:
return None if model_info is None else model_info.get("id")
def _retry_breadcrumbs(metadata: object) -> tuple[Mapping[str, object], ...]:
fields: Final = as_str_mapping(metadata)
breadcrumbs: Final = None if fields is None else fields.get("previous_models")
if breadcrumbs is None:
return ()
try:
return _RETRY_BREADCRUMBS_ADAPTER.validate_python(breadcrumbs)
except ValidationError:
return ()
def _deployment_ids_retried_in_request(request_kwargs: Mapping[str, object]) -> tuple[str, ...]:
breadcrumbs: Final = (
*_retry_breadcrumbs(request_kwargs.get("metadata")),
*_retry_breadcrumbs(request_kwargs.get("litellm_metadata")),
)
return tuple(
deployment_id for breadcrumb in breadcrumbs if isinstance(deployment_id := breadcrumb.get("deployment_id"), str)
)
def prefer_web_search_deployments(
deployments: list[dict[str, object]], # mutable-ok: router deployment-list contract
deployments: list[_DeploymentT], # mutable-ok: router deployment-list contract
request_kwargs: Mapping[str, object] | None,
) -> list[dict[str, object]]: # mutable-ok: returns a mutable deployment list
) -> list[_DeploymentT]: # mutable-ok: returns a mutable deployment list
if request_kwargs is None or not _is_web_search_request(
request_kwargs.get("tools"), request_kwargs.get("web_search_options")
):
@ -505,10 +531,11 @@ def prefer_web_search_deployments(
(
*as_excluded_deployment_ids(request_kwargs.get("_excluded_deployment_ids")),
*as_retry_skipped_deployment_ids(request_kwargs.get("_retry_skipped_deployment_ids")),
*_deployment_ids_retried_in_request(request_kwargs),
)
)
untried: Final = (d for d in deployments if _deployment_id(d) not in tried)
orders_with_an_untried_capable_deployment: Final = frozenset(
orders_with_an_untried_capable_deployment: Final = tuple(
litellm.utils.get_deployment_order(d) for d in untried if _deployment_supports_web_search(d)
)
if not orders_with_an_untried_capable_deployment:
@ -522,6 +549,36 @@ def prefer_web_search_deployments(
]
def pick_preferring_web_search(
pick: Callable[[list[_DeploymentT]], _Picked | None], # mutable-ok: router deployment-list contract
deployments: list[_DeploymentT], # mutable-ok: router deployment-list contract
request_kwargs: Mapping[str, object] | None,
) -> _Picked | None:
preferred: Final = prefer_web_search_deployments(deployments, request_kwargs)
if len(preferred) == len(deployments):
return pick(deployments)
try:
from_preferred: Final = pick(preferred)
except (litellm.RateLimitError, ValueError):
return pick(deployments)
return from_preferred if from_preferred is not None else pick(deployments)
async def apick_preferring_web_search(
pick: Callable[[list[_DeploymentT]], Awaitable[_Picked | None]], # mutable-ok: router deployment-list contract
deployments: list[_DeploymentT], # mutable-ok: router deployment-list contract
request_kwargs: Mapping[str, object] | None,
) -> _Picked | None:
preferred: Final = prefer_web_search_deployments(deployments, request_kwargs)
if len(preferred) == len(deployments):
return await pick(deployments)
try:
from_preferred: Final = await pick(preferred)
except (litellm.RateLimitError, ValueError):
return await pick(deployments)
return from_preferred if from_preferred is not None else await pick(deployments)
# Credential params that only one provider family reads, paired with the providers
# that read them. A deployment carrying them while resolving elsewhere is almost
# always a missing route prefix: `model: claude-sonnet-5` with `aws_region_name`

View file

@ -1,3 +1,4 @@
import asyncio
import logging
from itertools import chain
from typing import Dict, Final, List, Optional, Union
@ -15,6 +16,7 @@ from litellm.router_utils.common_utils import (
add_model_file_id_mappings,
filter_team_based_models,
filter_web_search_deployments,
pick_preferring_web_search,
prefer_web_search_deployments,
format_content_policy_no_fallback_note,
format_fallback_outcome_message,
@ -523,8 +525,11 @@ def test_sync_pick_keeps_a_group_without_web_search_for_openai_web_search(routin
@pytest.mark.asyncio
async def test_legacy_usage_based_async_pick_skips_deployments_without_web_search() -> None:
router: Final = _router_with_a_deployment_that_disables_web_search("usage-based-routing")
@pytest.mark.parametrize(
"routing_strategy", ["usage-based-routing", "usage-based-routing-v2", "least-busy", "latency-based-routing"]
)
async def test_async_strategy_pick_skips_deployments_without_web_search(routing_strategy: str) -> None:
router: Final = _router_with_a_deployment_that_disables_web_search(routing_strategy)
picks: Final = {
(
await router.async_get_available_deployment(
@ -706,6 +711,151 @@ async def test_web_search_request_prefers_the_capable_deployment_inside_its_orde
assert picks == {"search"}
@pytest.mark.asyncio
@pytest.mark.parametrize("metadata_field", ["metadata", "litellm_metadata"])
async def test_web_search_preference_gives_way_once_the_capable_deployment_failed_this_request(
metadata_field: str,
) -> None:
router: Final = _router_with_a_mixed_web_search_group(False)
picks: Final = {
(
await router.async_get_available_deployment(
model="mixed",
messages=[{"role": "user", "content": "hi"}],
request_kwargs={
**_ANTHROPIC_WEB_SEARCH_REQUEST,
metadata_field: {"previous_models": ({"deployment_id": "search", "exception_type": "Timeout"},)},
},
)
)["model_info"]["id"]
for _ in range(40)
}
assert picks == {"search", "no-search"}
@pytest.mark.asyncio
@_WEB_SEARCH_REQUESTS
async def test_retry_after_a_retryable_failure_on_the_capable_deployment_reaches_the_other_one(
web_search_request: dict[str, object],
) -> None:
router: Final = Router(
model_list=[
{
"model_name": "mixed",
"litellm_params": {
"model": "openai/gpt-x",
"api_key": "fake-key",
"mock_response": Exception("connection error"),
"weight": 1,
},
"model_info": {"id": "search"},
},
{
"model_name": "mixed",
"litellm_params": {"model": "openai/gpt-y", "api_key": "fake-key", "mock_response": "hi", "weight": 10**6},
"model_info": {"id": "no-search", "supports_web_search": False},
},
],
num_retries=1,
disable_cooldowns=True,
)
response: Final = await router.acompletion(
model="mixed", messages=[{"role": "user", "content": "hi"}], **web_search_request
)
assert response._hidden_params["model_id"] == "no-search"
def _rate_limited_web_search_group(routing_strategy: str) -> Router:
return Router(
model_list=[
{
"model_name": "mixed",
"litellm_params": {"model": "openai/gpt-x", "api_key": "fake-key", "mock_response": "hi", "rpm": 1},
"model_info": {"id": "search"},
},
{
"model_name": "mixed",
"litellm_params": {"model": "openai/gpt-y", "api_key": "fake-key", "mock_response": "hi", "rpm": 1000},
"model_info": {"id": "no-search", "supports_web_search": False},
},
],
routing_strategy=routing_strategy,
)
@pytest.mark.asyncio
@pytest.mark.parametrize("routing_strategy", ["usage-based-routing-v2", "cost-based-routing"])
async def test_web_search_request_goes_to_the_other_deployment_once_the_capable_one_is_at_its_rpm_limit(
routing_strategy: str,
) -> None:
router: Final = _rate_limited_web_search_group(routing_strategy)
served_by: Final = [] # mutable-ok: collects one id per sequential request
for _ in range(3):
response = await router.acompletion(
model="mixed", messages=[{"role": "user", "content": "hi"}], **_ANTHROPIC_WEB_SEARCH_REQUEST
)
served_by.append(response._hidden_params["model_id"])
await asyncio.sleep(0.2)
assert served_by == ["search", "no-search", "no-search"]
def test_sync_web_search_request_goes_to_the_other_deployment_once_the_capable_one_is_at_its_rpm_limit() -> None:
router: Final = _rate_limited_web_search_group("usage-based-routing-v2")
served_by: Final = tuple(
router.completion(
model="mixed", messages=[{"role": "user", "content": "hi"}], **_ANTHROPIC_WEB_SEARCH_REQUEST
)._hidden_params["model_id"]
for _ in range(3)
)
assert served_by == ("search", "no-search", "no-search")
def test_sync_pick_falls_back_to_every_deployment_when_the_strategy_finds_none_that_can_search() -> None:
deployments: Final = [
{"model_info": {"id": "search"}},
{"model_info": {"id": "no-search", "supports_web_search": False}},
]
picked: Final = pick_preferring_web_search(
lambda candidates: next((d for d in candidates if d["model_info"]["id"] == "no-search"), None),
deployments,
_ANTHROPIC_WEB_SEARCH_REQUEST,
)
assert picked == deployments[1]
@pytest.mark.asyncio
@pytest.mark.parametrize(
("search_order", "no_search_order"),
[([1], [1]), ({"tier": 1}, None)],
ids=["list-order", "dict-order"],
)
async def test_web_search_request_routes_on_a_group_whose_order_is_not_a_number(
search_order: object, no_search_order: object
) -> None:
router: Final = Router(
model_list=[
{
"model_name": "mixed",
"litellm_params": {"model": "openai/gpt-x", "api_key": "fake-key"},
"model_info": {"id": "search", "order": search_order},
},
{
"model_name": "mixed",
"litellm_params": {"model": "openai/gpt-y", "api_key": "fake-key"},
"model_info": {
"id": "no-search",
"supports_web_search": False,
**({} if no_search_order is None else {"order": no_search_order}),
},
},
],
)
deployment: Final = await router.async_get_available_deployment(
model="mixed", messages=[{"role": "user", "content": "hi"}], request_kwargs=dict(_ANTHROPIC_WEB_SEARCH_REQUEST)
)
assert deployment["model_info"]["id"] == "search"
@pytest.mark.asyncio
async def test_tag_routing_to_a_deployment_without_web_search_keeps_it() -> None:
deployment: Final = await _router_with_a_mixed_web_search_group(True).async_get_available_deployment(