fallback function support

packages/openassistants-fastapi/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "definitive-openassistants-fastapi"
-version = "0.0.10"
+version = "0.0.11"
 description = ""
 authors = ["Rick Lamers <[email protected]>"]
 readme = "README.md"

packages/openassistants/openassistants/functions/base.py

@@ -71,14 +71,13 @@ def get_signature(self) -> str:
         # convert JSON Schema types to Python types signature
         params_repr = PyRepr.repr_json_schema(self.get_parameters_json_schema())
 
-        sample_question_text = "\n".join(f"* {q}" for q in self.get_sample_questions())
-
         documentation = f"""\
 {self.get_description()}
-Example Questions:
-{sample_question_text}
 """
 
+        if len(self.get_sample_questions()) > 0:
+            documentation += "\n".join(f"* {q}" for q in self.get_sample_questions())
+
         # Construct the function signature
         signature = f"""\
 def {self.get_id()}({params_repr}) -> pd.DataFrame:
@@ -92,6 +91,13 @@ def {self.get_id()}({params_repr}) -> pd.DataFrame:
     async def get_entity_configs(self) -> Mapping[str, IEntityConfig]:
         pass
 
+    @abc.abstractmethod
+    def get_is_fallback(self) -> bool:
+        """
+        Whether this function is a fallback function.
+        Fallback functions are functions that are always available to be chosen if no other non-fallback functions are relevant.
+        """  # noqa: E501
+
     @abc.abstractmethod
     async def execute(
         self,
@@ -133,6 +139,7 @@ class BaseFunction(IFunction, BaseModel, abc.ABC):
     sample_questions: List[str] = []
     confirm: bool = False
     parameters: BaseFunctionParameters = BaseFunctionParameters()
+    is_fallback: bool = False
 
     def get_id(self) -> str:
         return self.id
@@ -158,6 +165,9 @@ def get_parameters_json_schema(self) -> JSONSchema:
     async def get_entity_configs(self) -> Mapping[str, IEntityConfig]:
         return {}
 
+    def get_is_fallback(self) -> bool:
+        return self.is_fallback
+
 
 class IFunctionLibrary(abc.ABC):
     @abc.abstractmethod

packages/openassistants/openassistants/llm_function_calling/selection.py

@@ -17,8 +17,13 @@ async def filter_functions(
     functions_text = "\n".join([f.get_signature() for f in functions])
     json_schema = {
         "type": "object",
-        "properties": {"function_name": {"type": "string"}},
+        "properties": {"function_name": {"$ref": "#/definitions/functions"}},
         "required": ["function_name"],
+        "definitions": {
+            "functions": {
+                "enum": [f.get_id() for f in functions],
+            }
+        },
     }
     messages = [
         HumanMessage(
@@ -54,7 +59,10 @@ async def select_function(
     user_query: str,
     chunk_size: int = 4,
 ) -> SelectFunctionResult:
-    subsets = chunk_list_by_max_size(functions, chunk_size)
+    non_fallbacks = [f for f in functions if not f.get_is_fallback()]
+    fallbacks = [f for f in functions if f.get_is_fallback()]
+
+    subsets = chunk_list_by_max_size(non_fallbacks, chunk_size)
 
     # Make LLM calls in parallel
     tasks = [
@@ -77,21 +85,39 @@ async def select_function(
     json_schema = {
         "type": "object",
         "properties": {
-            "function_name": {"type": "string"},
-            "suggested_function_names": {"type": "array", "items": {"type": "string"}},
+            "reason": {"type": "string"},
+            "function_name": {"$ref": "#/definitions/functions"},
+            "related_function_names": {
+                "type": "array",
+                "items": {"$ref": "#/definitions/functions"},
+            },
+        },
+        "definitions": {
+            "functions": {
+                "enum": [f.get_id() for f in selected_functions + fallbacks],
+            }
         },
     }
 
+    fallbacks_signatures = "\n".join([f.get_signature() for f in fallbacks])
+
     selection_messages = [
         HumanMessage(
-            content=f"""Prior selection reduced the candidates to these functions:
+            content=f"""\
+Prior selection reduced the candidates to these functions:
+
 {selected_functions_signatures}
 
+These fallback functions can be used when none of the above functions are a good match:
+
+{fallbacks_signatures}
+
 Scenario 1: There is a function in the list of candidates that is a match to the user query.
 Action: provide the name of the function as the 'function_name' argument.
 
 Scenario 2: None of the functions in the list of candidates match the user query.
-Action: select related functions from the list of candidates as the 'suggested_function_names' argument. You are also allowed to return an empty list of suggested functions if you think none of the functions are a good match.
+Action: select related functions from the list of candidates as the 'related_function_names' argument.
+You are also allowed to return an empty list of related functions if you think none of the functions are a good match.
 
 First decide which of the two scenarios is the case. Then take the appropriate action.
 
@@ -114,10 +140,12 @@ async def select_function(
     suggested_function_names = json_result.get("related_function_names", [])
 
     selected_function = next(
-        (f for f in selected_functions if f.get_id() == function_name), None
+        (f for f in selected_functions + fallbacks if f.get_id() == function_name), None
     )
     suggested_functions = [
-        f for f in selected_functions if f.get_id() in suggested_function_names
+        f
+        for f in selected_functions + fallbacks
+        if f.get_id() in suggested_function_names
     ] or None
 
     return SelectFunctionResult(

packages/openassistants/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "definitive-openassistants"
-version = "0.0.10"
+version = "0.0.11"
 description = ""
 authors = ["Rick Lamers <[email protected]>"]
 readme = "README.md"