Movatterモバイル変換

[0]ホーム
URL:

Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

 Sign up
Appearance settings

Commit22eb7e5

Browse files
GWealecopybara-github
authored andcommitted
feat: Add support for parsing inline JSON tool calls in LiteLLM responses
Close#1968Co-authored-by: George Weale <gweale@google.com>PiperOrigin-RevId: 831911719
1 parent2efc184 commit22eb7e5

File tree

4 files changed

+511
-53
lines changed

4 files changed

+511
-53
lines changed
Lines changed: 17 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,17 @@
1+
# Copyright 2025 Google LLC
2+
#
3+
# Licensed under the Apache License, Version 2.0 (the "License");
4+
# you may not use this file except in compliance with the License.
5+
# You may obtain a copy of the License at
6+
#
7+
# http://www.apache.org/licenses/LICENSE-2.0
8+
#
9+
# Unless required by applicable law or agreed to in writing, software
10+
# distributed under the License is distributed on an "AS IS" BASIS,
11+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+
# See the License for the specific language governing permissions and
13+
# limitations under the License.
14+
15+
from __future__importannotations
16+
17+
from .importagent
Lines changed: 174 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,174 @@
1+
# Copyright 2025 Google LLC
2+
#
3+
# Licensed under the Apache License, Version 2.0 (the "License");
4+
# you may not use this file except in compliance with the License.
5+
# You may obtain a copy of the License at
6+
#
7+
# http://www.apache.org/licenses/LICENSE-2.0
8+
#
9+
# Unless required by applicable law or agreed to in writing, software
10+
# distributed under the License is distributed on an "AS IS" BASIS,
11+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+
# See the License for the specific language governing permissions and
13+
# limitations under the License.
14+
15+
from __future__importannotations
16+
17+
importdatetime
18+
importjson
19+
importre
20+
fromtypingimportAny
21+
fromzoneinfoimportZoneInfo
22+
fromzoneinfoimportZoneInfoNotFoundError
23+
24+
fromgoogle.adk.agents.llm_agentimportAgent
25+
fromgoogle.adk.models.lite_llmimportLiteLlm
26+
fromgoogle.adk.models.lite_llmimportLiteLLMClient
27+
28+
29+
classInlineJsonToolClient(LiteLLMClient):
30+
"""LiteLLM client that emits inline JSON tool calls for testing."""
31+
32+
asyncdefacompletion(self,model,messages,tools,**kwargs):
33+
deltools,kwargs# Only needed for API parity.
34+
35+
tool_message=_find_last_role(messages,role="tool")
36+
iftool_message:
37+
tool_summary=_coerce_to_text(tool_message.get("content"))
38+
return {
39+
"id":"mock-inline-tool-final-response",
40+
"model":model,
41+
"choices": [{
42+
"message": {
43+
"role":"assistant",
44+
"content": (
45+
f"The instrumentation tool responded with:{tool_summary}"
46+
),
47+
},
48+
"finish_reason":"stop",
49+
}],
50+
"usage": {
51+
"prompt_tokens":60,
52+
"completion_tokens":12,
53+
"total_tokens":72,
54+
},
55+
}
56+
57+
timezone=_extract_timezone(messages)or"Asia/Taipei"
58+
inline_call=json.dumps(
59+
{
60+
"name":"get_current_time",
61+
"arguments": {"timezone_str":timezone},
62+
},
63+
separators=(",",":"),
64+
)
65+
66+
return {
67+
"id":"mock-inline-tool-call",
68+
"model":model,
69+
"choices": [{
70+
"message": {
71+
"role":"assistant",
72+
"content": (
73+
f"{inline_call}\nLet me double-check the clock for you."
74+
),
75+
},
76+
"finish_reason":"tool_calls",
77+
}],
78+
"usage": {
79+
"prompt_tokens":45,
80+
"completion_tokens":15,
81+
"total_tokens":60,
82+
},
83+
}
84+
85+
86+
def_find_last_role(
87+
messages:list[dict[str,Any]],role:str
88+
)->dict[str,Any]:
89+
"""Returns the last message with the given role."""
90+
formessageinreversed(messages):
91+
ifmessage.get("role")==role:
92+
returnmessage
93+
return {}
94+
95+
96+
def_coerce_to_text(content:Any)->str:
97+
"""Best-effort conversion from OpenAI message content to text."""
98+
ifisinstance(content,str):
99+
returncontent
100+
ifisinstance(content,dict):
101+
return_coerce_to_text(content.get("text"))
102+
ifisinstance(content,list):
103+
texts= []
104+
forpartincontent:
105+
ifisinstance(part,dict):
106+
texts.append(part.get("text")or"")
107+
elifisinstance(part,str):
108+
texts.append(part)
109+
return" ".join(textfortextintextsiftext)
110+
return""
111+
112+
113+
_TIMEZONE_PATTERN=re.compile(r"([A-Za-z]+/[A-Za-z_]+)")
114+
115+
116+
def_extract_timezone(messages:list[dict[str,Any]])->str|None:
117+
"""Extracts an IANA timezone string from the last user message."""
118+
user_message=_find_last_role(messages,role="user")
119+
text=_coerce_to_text(user_message.get("content"))
120+
ifnottext:
121+
returnNone
122+
match=_TIMEZONE_PATTERN.search(text)
123+
ifmatch:
124+
returnmatch.group(1)
125+
lowered=text.lower()
126+
if"taipei"inlowered:
127+
return"Asia/Taipei"
128+
if"new york"inlowered:
129+
return"America/New_York"
130+
if"london"inlowered:
131+
return"Europe/London"
132+
if"tokyo"inlowered:
133+
return"Asia/Tokyo"
134+
returnNone
135+
136+
137+
defget_current_time(timezone_str:str)->dict[str,str]:
138+
"""Returns mock current time for the provided timezone."""
139+
try:
140+
tz=ZoneInfo(timezone_str)
141+
exceptZoneInfoNotFoundErrorasexc:
142+
return {
143+
"status":"error",
144+
"report":f"Unable to parse timezone '{timezone_str}':{exc}",
145+
}
146+
now=datetime.datetime.now(tz)
147+
return {
148+
"status":"success",
149+
"report": (
150+
f"The current time in{timezone_str} is"
151+
f"{now.strftime('%Y-%m-%d %H:%M:%S %Z')}."
152+
),
153+
}
154+
155+
156+
_mock_model=LiteLlm(
157+
model="mock/inline-json-tool-calls",
158+
llm_client=InlineJsonToolClient(),
159+
)
160+
161+
root_agent=Agent(
162+
name="litellm_inline_tool_tester",
163+
model=_mock_model,
164+
description=(
165+
"Demonstrates LiteLLM inline JSON tool-call parsing without an external"
166+
" VLLM deployment."
167+
),
168+
instruction=(
169+
"You are a deterministic clock assistant. Always call the"
170+
" get_current_time tool before answering user questions. After the tool"
171+
" responds, summarize what it returned."
172+
),
173+
tools=[get_current_time],
174+
)

‎src/google/adk/models/lite_llm.py‎

Lines changed: 134 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -31,6 +31,7 @@
3131
fromtypingimportTuple
3232
fromtypingimportTypedDict
3333
fromtypingimportUnion
34+
importuuid
3435
importwarnings
3536

3637
fromgoogle.genaiimporttypes
@@ -64,6 +65,7 @@
6465
_NEW_LINE="\n"
6566
_EXCLUDED_PART_FIELD= {"inline_data": {"data"}}
6667
_LITELLM_STRUCTURED_TYPES= {"json_object","json_schema"}
68+
_JSON_DECODER=json.JSONDecoder()
6769

6870
# Mapping of LiteLLM finish_reason strings to FinishReason enum values
6971
# Note: tool_calls/function_call map to STOP because:
@@ -431,6 +433,118 @@ def _get_content(
431433
returncontent_objects
432434

433435

436+
def_build_tool_call_from_json_dict(
437+
candidate:Any,*,index:int
438+
)->Optional[ChatCompletionMessageToolCall]:
439+
"""Creates a tool call object from JSON content embedded in text."""
440+
441+
ifnotisinstance(candidate,dict):
442+
returnNone
443+
444+
name=candidate.get("name")
445+
args=candidate.get("arguments")
446+
ifnotisinstance(name,str)orargsisNone:
447+
returnNone
448+
449+
ifisinstance(args,str):
450+
arguments_payload=args
451+
else:
452+
try:
453+
arguments_payload=json.dumps(args,ensure_ascii=False)
454+
except (TypeError,ValueError):
455+
arguments_payload=_safe_json_serialize(args)
456+
457+
call_id=candidate.get("id")orf"adk_tool_call_{uuid.uuid4().hex}"
458+
call_index=candidate.get("index")
459+
ifisinstance(call_index,int):
460+
index=call_index
461+
462+
function=Function(
463+
name=name,
464+
arguments=arguments_payload,
465+
)
466+
# Some LiteLLM types carry an `index` field only in streaming contexts,
467+
# so guard the assignment to stay compatible with older versions.
468+
ifhasattr(function,"index"):
469+
function.index=index# type: ignore[attr-defined]
470+
471+
tool_call=ChatCompletionMessageToolCall(
472+
type="function",
473+
id=str(call_id),
474+
function=function,
475+
)
476+
# Same reasoning as above: not every ChatCompletionMessageToolCall exposes it.
477+
ifhasattr(tool_call,"index"):
478+
tool_call.index=index# type: ignore[attr-defined]
479+
480+
returntool_call
481+
482+
483+
def_parse_tool_calls_from_text(
484+
text_block:str,
485+
)->tuple[list[ChatCompletionMessageToolCall],Optional[str]]:
486+
"""Extracts inline JSON tool calls from LiteLLM text responses."""
487+
488+
tool_calls= []
489+
ifnottext_block:
490+
returntool_calls,None
491+
492+
remainder_segments= []
493+
cursor=0
494+
text_length=len(text_block)
495+
496+
whilecursor<text_length:
497+
brace_index=text_block.find("{",cursor)
498+
ifbrace_index==-1:
499+
remainder_segments.append(text_block[cursor:])
500+
break
501+
502+
remainder_segments.append(text_block[cursor:brace_index])
503+
try:
504+
candidate,end=_JSON_DECODER.raw_decode(text_block,brace_index)
505+
exceptjson.JSONDecodeError:
506+
remainder_segments.append(text_block[brace_index])
507+
cursor=brace_index+1
508+
continue
509+
510+
tool_call=_build_tool_call_from_json_dict(
511+
candidate,index=len(tool_calls)
512+
)
513+
iftool_call:
514+
tool_calls.append(tool_call)
515+
else:
516+
remainder_segments.append(text_block[brace_index:end])
517+
cursor=end
518+
519+
remainder="".join(segmentforsegmentinremainder_segmentsifsegment)
520+
remainder=remainder.strip()
521+
522+
returntool_calls,remainderorNone
523+
524+
525+
def_split_message_content_and_tool_calls(
526+
message:Message,
527+
)->tuple[Optional[OpenAIMessageContent],list[ChatCompletionMessageToolCall]]:
528+
"""Returns message content and tool calls, parsing inline JSON when needed."""
529+
530+
existing_tool_calls=message.get("tool_calls")or []
531+
normalized_tool_calls= (
532+
list(existing_tool_calls)ifexisting_tool_callselse []
533+
)
534+
content=message.get("content")
535+
536+
# LiteLLM responses either provide structured tool_calls or inline JSON, not
537+
# both. When tool_calls are present we trust them and skip the fallback parser.
538+
ifnormalized_tool_callsornotisinstance(content,str):
539+
returncontent,normalized_tool_calls
540+
541+
fallback_tool_calls,remainder=_parse_tool_calls_from_text(content)
542+
iffallback_tool_calls:
543+
returnremainder,fallback_tool_calls
544+
545+
returncontent, []
546+
547+
434548
def_to_litellm_role(role:Optional[str])->Literal["user","assistant"]:
435549
"""Converts a types.Content role to a litellm role.
436550
@@ -584,15 +698,24 @@ def _model_response_to_chunk(
584698
ifmessageisNoneandresponse["choices"][0].get("delta",None):
585699
message=response["choices"][0]["delta"]
586700

587-
ifmessage.get("content",None):
588-
yieldTextChunk(text=message.get("content")),finish_reason
701+
message_content:Optional[OpenAIMessageContent]=None
702+
tool_calls:list[ChatCompletionMessageToolCall]= []
703+
ifmessageisnotNone:
704+
(
705+
message_content,
706+
tool_calls,
707+
)=_split_message_content_and_tool_calls(message)
589708

590-
ifmessage.get("tool_calls",None):
591-
fortool_callinmessage.get("tool_calls"):
709+
ifmessage_content:
710+
yieldTextChunk(text=message_content),finish_reason
711+
712+
iftool_calls:
713+
foridx,tool_callinenumerate(tool_calls):
592714
# aggregate tool_call
593715
iftool_call.type=="function":
594716
func_name=tool_call.function.name
595717
func_args=tool_call.function.arguments
718+
func_index=getattr(tool_call,"index",idx)
596719

597720
# Ignore empty chunks that don't carry any information.
598721
ifnotfunc_nameandnotfunc_args:
@@ -602,12 +725,10 @@ def _model_response_to_chunk(
602725
id=tool_call.id,
603726
name=func_name,
604727
args=func_args,
605-
index=tool_call.index,
728+
index=func_index,
606729
),finish_reason
607730

608-
iffinish_reasonandnot (
609-
message.get("content",None)ormessage.get("tool_calls",None)
610-
):
731+
iffinish_reasonandnot (message_contentortool_calls):
611732
yieldNone,finish_reason
612733

613734
ifnotmessage:
@@ -687,11 +808,12 @@ def _message_to_generate_content_response(
687808
"""
688809

689810
parts= []
690-
ifmessage.get("content",None):
691-
parts.append(types.Part.from_text(text=message.get("content")))
811+
message_content,tool_calls=_split_message_content_and_tool_calls(message)
812+
ifisinstance(message_content,str)andmessage_content:
813+
parts.append(types.Part.from_text(text=message_content))
692814

693-
ifmessage.get("tool_calls",None):
694-
fortool_callinmessage.get("tool_calls"):
815+
iftool_calls:
816+
fortool_callintool_calls:
695817
iftool_call.type=="function":
696818
part=types.Part.from_function_call(
697819
name=tool_call.function.name,

0 commit comments

Comments
 (0)
[8]ページ先頭
©2009-2026 Movatter.jp