Merge branch 'develop'

README.md

@@ -498,7 +498,7 @@ Conversation options are represented as features. They can be configured from Ap
 | `recognition_retry_max` | The maximum number of retries for voice recognition. | `int` | 2 |
 | `recording_enabled` | Whether call recording is enabled. | `bool` | false |
 | `slow_llm_for_chat` | Whether to use the slow LLM for chat. | `bool` | false |
-| `vad_cutoff_timeout_ms` | The cutoff timeout for voice activity detection in secs. | `int` | 150 |
+| `vad_cutoff_timeout_ms` | The cutoff timeout for voice activity detection in secs. | `int` | 250 |
 | `vad_silence_timeout_ms` | The timeout for phone silence in secs. | `int` | 500 |
 | `vad_threshold` | The threshold for voice activity detection. | `float` | 0.5 |
 

app/helpers/call_llm.py

@@ -178,7 +178,7 @@ async def _stop_callback() -> None:
 
         async def _commit_answer(
             wait: bool,
-            tool_blacklist: set[str] | None = None,
+            tool_blacklist: set[str] = set(),
         ) -> None:
             """
             Process the response.
@@ -289,9 +289,9 @@ async def _continue_chat(  # noqa: PLR0915, PLR0913
     client: CallAutomationClient,
     post_callback: Callable[[CallStateModel], Awaitable[None]],
     scheduler: Scheduler,
-    tool_blacklist: set[str] | None,
     training_callback: Callable[[CallStateModel], Awaitable[None]],
     tts_client: SpeechSynthesizer,
+    tool_blacklist: set[str] = set(),
     _iterations_remaining: int = 3,
 ) -> CallStateModel:
     """
@@ -482,7 +482,7 @@ async def _generate_chat_completion(  # noqa: PLR0913, PLR0911, PLR0912, PLR0915
     client: CallAutomationClient,
     post_callback: Callable[[CallStateModel], Awaitable[None]],
     scheduler: Scheduler,
-    tool_blacklist: set[str] | None,
+    tool_blacklist: set[str],
     tts_callback: Callable[[str, MessageStyleEnum], Awaitable[None]],
     tts_client: SpeechSynthesizer,
     use_tools: bool,
@@ -539,7 +539,7 @@ async def _content_callback(buffer: str) -> None:
     if not use_tools:
         logger.warning("Tools disabled for this chat")
     else:
-        tools = await plugins.to_openai(tool_blacklist)
+        tools = await plugins.to_openai(frozenset(tool_blacklist))
         # logger.debug("Tools: %s", tools)
 
     # Execute LLM inference
@@ -626,7 +626,13 @@ async def _content_callback(buffer: str) -> None:
         scheduler=scheduler,
     ):
         await asyncio.gather(
-            *[plugins.execute_tool(tool_call) for tool_call in tool_calls]
+            *[
+                plugins.execute(
+                    blacklist=tool_blacklist,
+                    tool=tool_call,
+                )
+                for tool_call in tool_calls
+            ]
         )
 
     # Update call model if object reference changed

app/helpers/config_models/prompts.py

@@ -62,7 +62,8 @@ class LlmModel(BaseModel):
         - Provide a clear and concise summary of the conversation at the beginning of each call
         - Respond only if it is related to the objective or the claim
         - To list things, use bullet points or numbered lists
-        - Use a lot of discourse markers, fillers, to make the conversation human-like
+        - Use a lot of discourse markers, fillers, to make the conversation human-like (e.g., "Well, let me think...", "So, what I can do for you is...", "I see, you are in Paris...")
+        - Use short sentences and simple words
         - Use tools as often as possible and describe the actions you take
         - When the customer says a word and then spells out letters, this means that the word is written in the way the customer spelled it (e.g., "I live in Paris PARIS" -> "Paris", "My name is John JOHN" -> "John", "My email is Clemence CLEMENCE at gmail dot com" -> "[email protected]")
         - Work for {bot_company}, not someone else
@@ -147,7 +148,7 @@ class LlmModel(BaseModel):
         Conversation objective: Fill the claim with the customer. Claim is about a car accident.
         User: action=talk I had an accident this morning, I was shopping. Let me send the exact location by SMS.
         User: action=sms At the corner of Rue de la Paix and Rue de Rivoli.
-        Tools: update incident location,n
+        Tools: update incident location
         Assistant: style=sad I get it, you had an accident this morning while shopping. style=none I have updated your file with the location you sent me by SMS. style=cheerful Is it correct?
 
         ## Example 7

app/helpers/features.py

@@ -72,7 +72,7 @@ async def vad_silence_timeout_ms(scheduler: Scheduler) -> int:
 
 async def vad_cutoff_timeout_ms(scheduler: Scheduler) -> int:
     return await _default(
-        default=150,
+        default=250,
         key="vad_cutoff_timeout_ms",
         scheduler=scheduler,
         type_res=int,