|
25 | 25 |
|
26 | 26 |
|
27 | 27 | class GeminiLlmConnection(BaseLlmConnection):
|
28 |
| - """The Gemini model connection.""" |
29 |
| - |
30 |
| - def __init__(self, gemini_session: live.AsyncSession): |
31 |
| - self._gemini_session = gemini_session |
32 |
| - |
33 |
| - async def send_history(self, history: list[types.Content]): |
34 |
| - """Sends the conversation history to the gemini model. |
35 |
| -
|
36 |
| - You call this method right after setting up the model connection. |
37 |
| - The model will respond if the last content is from user, otherwise it will |
38 |
| - wait for new user input before responding. |
39 |
| -
|
40 |
| - Args: |
41 |
| - history: The conversation history to send to the model. |
42 |
| - """ |
43 |
| - |
44 |
| - # TODO: Remove this filter and translate unary contents to streaming |
45 |
| - # contents properly. |
46 |
| - |
47 |
| - # We ignore any audio from user during the agent transfer phase |
48 |
| - contents = [ |
49 |
| - content for content in history if content.parts and content.parts[0].text |
50 |
| - ] |
51 |
| - |
52 |
| - if contents: |
53 |
| - await self._gemini_session.send( |
54 |
| - input=types.LiveClientContent( |
55 |
| - turns=contents, |
56 |
| - turn_complete=contents[-1].role == "user", |
57 |
| - ), |
58 |
| - ) |
59 |
| - else: |
60 |
| - logger.info("no content is sent") |
61 |
| - |
62 |
| - async def send_content(self, content: types.Content): |
63 |
| - """Sends a user content to the gemini model. |
64 |
| -
|
65 |
| - The model will respond immediately upon receiving the content. |
66 |
| - If you send function responses, all parts in the content should be function |
67 |
| - responses. |
68 |
| -
|
69 |
| - Args: |
70 |
| - content: The content to send to the model. |
71 |
| - """ |
72 |
| - |
73 |
| - assert content.parts |
74 |
| - if content.parts[0].function_response: |
75 |
| - # All parts have to be function responses. |
76 |
| - function_responses = [part.function_response for part in content.parts] |
77 |
| - logger.debug("Sending LLM function response: %s", function_responses) |
78 |
| - await self._gemini_session.send( |
79 |
| - input=types.LiveClientToolResponse( |
80 |
| - function_responses=function_responses |
81 |
| - ), |
82 |
| - ) |
83 |
| - else: |
84 |
| - logger.debug("Sending LLM new content %s", content) |
85 |
| - await self._gemini_session.send( |
86 |
| - input=types.LiveClientContent( |
87 |
| - turns=[content], |
88 |
| - turn_complete=True, |
| 28 | + """The Gemini model connection.""" |
| 29 | + |
| 30 | + def __init__(self, gemini_session: live.AsyncSession): |
| 31 | + self._gemini_session = gemini_session |
| 32 | + |
| 33 | + async def send_history(self, history: list[types.Content]): |
| 34 | + """Sends the conversation history to the gemini model. |
| 35 | +
|
| 36 | + You call this method right after setting up the model connection. |
| 37 | + The model will respond if the last content is from user, otherwise it will |
| 38 | + wait for new user input before responding. |
| 39 | +
|
| 40 | + Args: |
| 41 | + history: The conversation history to send to the model. |
| 42 | + """ |
| 43 | + |
| 44 | + # TODO: Remove this filter and translate unary contents to streaming |
| 45 | + # contents properly. |
| 46 | + |
| 47 | + # We ignore any audio from user during the agent transfer phase |
| 48 | + contents = [ |
| 49 | + content |
| 50 | + for content in history |
| 51 | + if content.parts and content.parts[0].text |
| 52 | + ] |
| 53 | + |
| 54 | + if contents: |
| 55 | + await self._gemini_session.send( |
| 56 | + input=types.LiveClientContent( |
| 57 | + turns=contents, |
| 58 | + turn_complete=contents[-1].role == 'user', |
| 59 | + ), |
| 60 | + ) |
| 61 | + else: |
| 62 | + logger.info('no content is sent') |
| 63 | + |
| 64 | + async def send_content(self, content: types.Content): |
| 65 | + """Sends a user content to the gemini model. |
| 66 | +
|
| 67 | + The model will respond immediately upon receiving the content. |
| 68 | + If you send function responses, all parts in the content should be function |
| 69 | + responses. |
| 70 | +
|
| 71 | + Args: |
| 72 | + content: The content to send to the model. |
| 73 | + """ |
| 74 | + |
| 75 | + assert content.parts |
| 76 | + if content.parts[0].function_response: |
| 77 | + # All parts have to be function responses. |
| 78 | + function_responses = [part.function_response for part in content.parts] |
| 79 | + logger.debug('Sending LLM function response: %s', function_responses) |
| 80 | + await self._gemini_session.send( |
| 81 | + input=types.LiveClientToolResponse( |
| 82 | + function_responses=function_responses |
| 83 | + ), |
| 84 | + ) |
| 85 | + else: |
| 86 | + logger.debug('Sending LLM new content %s', content) |
| 87 | + await self._gemini_session.send( |
| 88 | + input=types.LiveClientContent( |
| 89 | + turns=[content], |
| 90 | + turn_complete=True, |
| 91 | + ) |
| 92 | + ) |
| 93 | + |
| 94 | + async def send_realtime(self, blob: types.Blob): |
| 95 | + """Sends a chunk of audio or a frame of video to the model in realtime. |
| 96 | +
|
| 97 | + Args: |
| 98 | + blob: The blob to send to the model. |
| 99 | + """ |
| 100 | + |
| 101 | + input_blob = blob.model_dump() |
| 102 | + logger.debug('Sending LLM Blob: %s', input_blob) |
| 103 | + await self._gemini_session.send(input=input_blob) |
| 104 | + |
| 105 | + def __build_full_text_response(self, text: str): |
| 106 | + """Builds a full text response. |
| 107 | +
|
| 108 | + The text should not partial and the returned LlmResponse is not be |
| 109 | + partial. |
| 110 | +
|
| 111 | + Args: |
| 112 | + text: The text to be included in the response. |
| 113 | +
|
| 114 | + Returns: |
| 115 | + An LlmResponse containing the full text. |
| 116 | + """ |
| 117 | + return LlmResponse( |
| 118 | + content=types.Content( |
| 119 | + role='model', |
| 120 | + parts=[types.Part.from_text(text=text)], |
| 121 | + ), |
| 122 | + ) |
| 123 | + |
| 124 | + async def receive(self) -> AsyncGenerator[LlmResponse, None]: |
| 125 | + """Receives the model response using the llm server connection. |
| 126 | +
|
| 127 | + Yields: |
| 128 | + LlmResponse: The model response. |
| 129 | + """ |
| 130 | + |
| 131 | + text = '' |
| 132 | + async for message in self._gemini_session.receive(): |
| 133 | + logger.debug('Got LLM Live message: %s', message) |
| 134 | + if message.server_content: |
| 135 | + content = message.server_content.model_turn |
| 136 | + if content and content.parts: |
| 137 | + llm_response = LlmResponse( |
| 138 | + content=content, interrupted=message.server_content.interrupted |
| 139 | + ) |
| 140 | + if content.parts[0].text: |
| 141 | + text += content.parts[0].text |
| 142 | + llm_response.partial = True |
| 143 | + # don't yield the merged text event when receiving audio data |
| 144 | + elif text and not content.parts[0].inline_data: |
| 145 | + yield self.__build_full_text_response(text) |
| 146 | + text = '' |
| 147 | + yield llm_response |
| 148 | + if ( |
| 149 | + message.server_content.input_transcription |
| 150 | + and message.server_content.input_transcription.text |
| 151 | + ): |
| 152 | + user_text = message.server_content.input_transcription.text |
| 153 | + parts = [ |
| 154 | + types.Part.from_text( |
| 155 | + text=user_text, |
89 | 156 | )
|
| 157 | + ] |
| 158 | + llm_response = LlmResponse( |
| 159 | + content=types.Content(role='user', parts=parts) |
90 | 160 | )
|
| 161 | + yield llm_response |
| 162 | + if ( |
| 163 | + message.server_content.output_transcription |
| 164 | + and message.server_content.output_transcription.text |
| 165 | + ): |
| 166 | + # TODO: Right now, we just support output_transcription without |
| 167 | + # changing interface and data protocol. Later, we can consider to |
| 168 | + # support output_transcription as a separate field in LlmResponse. |
| 169 | + |
| 170 | + # Transcription is always considered as partial event |
| 171 | + # We rely on other control signals to determine when to yield the |
| 172 | + # full text response(turn_complete, interrupted, or tool_call). |
| 173 | + text += message.server_content.output_transcription.text |
| 174 | + parts = [ |
| 175 | + types.Part.from_text( |
| 176 | + text=message.server_content.output_transcription.text |
| 177 | + ) |
| 178 | + ] |
| 179 | + llm_response = LlmResponse( |
| 180 | + content=types.Content(role='model', parts=parts), partial=True |
| 181 | + ) |
| 182 | + yield llm_response |
| 183 | + |
| 184 | + if message.server_content.turn_complete: |
| 185 | + if text: |
| 186 | + yield self.__build_full_text_response(text) |
| 187 | + text = '' |
| 188 | + yield LlmResponse( |
| 189 | + turn_complete=True, interrupted=message.server_content.interrupted |
| 190 | + ) |
| 191 | + break |
| 192 | + # in case of empty content or parts, we sill surface it |
| 193 | + # in case it's an interrupted message, we merge the previous partial |
| 194 | + # text. Other we don't merge. because content can be none when model |
| 195 | + # safety threshold is triggered |
| 196 | + if message.server_content.interrupted and text: |
| 197 | + yield self.__build_full_text_response(text) |
| 198 | + text = '' |
| 199 | + yield LlmResponse(interrupted=message.server_content.interrupted) |
| 200 | + if message.tool_call: |
| 201 | + if text: |
| 202 | + yield self.__build_full_text_response(text) |
| 203 | + text = '' |
| 204 | + parts = [ |
| 205 | + types.Part(function_call=function_call) |
| 206 | + for function_call in message.tool_call.function_calls |
| 207 | + ] |
| 208 | + yield LlmResponse(content=types.Content(role='model', parts=parts)) |
| 209 | + |
| 210 | + async def close(self): |
| 211 | + """Closes the llm server connection.""" |
91 | 212 |
|
92 |
| - async def send_realtime(self, blob: types.Blob): |
93 |
| - """Sends a chunk of audio or a frame of video to the model in realtime. |
94 |
| -
|
95 |
| - Args: |
96 |
| - blob: The blob to send to the model. |
97 |
| - """ |
98 |
| - |
99 |
| - input_blob = blob.model_dump() |
100 |
| - logger.debug("Sending LLM Blob: %s", input_blob) |
101 |
| - await self._gemini_session.send(input=input_blob) |
102 |
| - |
103 |
| - def __build_full_text_response(self, text: str): |
104 |
| - """Builds a full text response. |
105 |
| -
|
106 |
| - The text should not partial and the returned LlmResponse is not be |
107 |
| - partial. |
108 |
| -
|
109 |
| - Args: |
110 |
| - text: The text to be included in the response. |
111 |
| -
|
112 |
| - Returns: |
113 |
| - An LlmResponse containing the full text. |
114 |
| - """ |
115 |
| - return LlmResponse( |
116 |
| - content=types.Content( |
117 |
| - role="model", |
118 |
| - parts=[types.Part.from_text(text=text)], |
119 |
| - ), |
120 |
| - ) |
121 |
| - |
122 |
| - async def receive(self) -> AsyncGenerator[LlmResponse, None]: |
123 |
| - """Receives the model response using the llm server connection. |
124 |
| -
|
125 |
| - Yields: |
126 |
| - LlmResponse: The model response. |
127 |
| - """ |
128 |
| - |
129 |
| - text = "" |
130 |
| - async for message in self._gemini_session.receive(): |
131 |
| - logger.debug("Got LLM Live message: %s", message) |
132 |
| - |
133 |
| - if message.server_content: |
134 |
| - content = message.server_content.model_turn |
135 |
| - if content and content.parts: |
136 |
| - llm_response = LlmResponse( |
137 |
| - content=content, interrupted=message.server_content.interrupted |
138 |
| - ) |
139 |
| - if content.parts[0].text: |
140 |
| - text += content.parts[0].text |
141 |
| - llm_response.partial = True |
142 |
| - # don't yield the merged text event when receiving audio data |
143 |
| - elif text and not content.parts[0].inline_data: |
144 |
| - yield self.__build_full_text_response(text) |
145 |
| - text = "" |
146 |
| - yield llm_response |
147 |
| - if ( |
148 |
| - message.server_content.input_transcription |
149 |
| - and message.server_content.input_transcription.text |
150 |
| - ): |
151 |
| - user_text = message.server_content.input_transcription.text |
152 |
| - parts = [ |
153 |
| - types.Part.from_text( |
154 |
| - text=user_text, |
155 |
| - ) |
156 |
| - ] |
157 |
| - llm_response = LlmResponse( |
158 |
| - content=types.Content(role="user", parts=parts) |
159 |
| - ) |
160 |
| - yield llm_response |
161 |
| - if ( |
162 |
| -
10000
message.server_content.output_transcription |
163 |
| - and message.server_content.output_transcription.text |
164 |
| - ): |
165 |
| - # TODO: Right now, we just support output_transcription without |
166 |
| - # changing interface and data protocol. Later, we can consider to |
167 |
| - # support output_transcription as a separate field in LlmResponse. |
168 |
| - |
169 |
| - # Transcription is always considered as partial event |
170 |
| - # We rely on other control signals to determine when to yield the |
171 |
| - # full text response(turn_complete, interrupted, or tool_call). |
172 |
| - text += message.server_content.output_transcription.text |
173 |
| - parts = [ |
174 |
| - types.Part.from_text( |
175 |
| - text=message.server_content.output_transcription.text |
176 |
| - ) |
177 |
| - ] |
178 |
| - llm_response = LlmResponse( |
179 |
| - content=types.Content(role="model", parts=parts), partial=True |
180 |
| - ) |
181 |
| - yield llm_response |
182 |
| - |
183 |
| - if message.server_content.turn_complete: |
184 |
| - if text: |
185 |
| - yield self.__build_full_text_response(text) |
186 |
| - text = "" |
187 |
| - yield LlmResponse( |
188 |
| - turn_complete=True, |
189 |
| - interrupted=message.server_content.interrupted, |
190 |
| - ) |
191 |
| - break |
192 |
| - # in case of empty content or parts, we sill surface it |
193 |
| - # in case it's an interrupted message, we merge the previous partial |
194 |
| - # text. Other we don't merge. because content can be none when model |
195 |
| - # safety threshold is triggered |
196 |
| - if message.server_content.interrupted and text: |
197 |
| - yield self.__build_full_text_response(text) |
198 |
| - text = "" |
199 |
| - yield LlmResponse(interrupted=message.server_content.interrupted) |
200 |
| - if message.tool_call: |
201 |
| - if text: |
202 |
| - yield self.__build_full_text_response(text) |
203 |
| - text = "" |
204 |
| - parts = [ |
205 |
| - types.Part(function_call=function_call) |
206 |
| - for function_call in message.tool_call.function_calls |
207 |
| - ] |
208 |
| - yield LlmResponse(content=types.Content(role="model", parts=parts)) |
209 |
| - else: |
210 |
| - pass |
211 |
| - |
212 |
| - async def close(self): |
213 |
| - """Closes the llm server connection.""" |
214 |
| - |
215 |
| - await self._gemini_session.close() |
| 213 | + await self._gemini_session.close() |
0 commit comments