more progress

Author: dtam

guardrails/actions/reask.py

@@ -284,10 +284,7 @@ def get_reask_setup_for_string(
     )
 
     instructions = None
-    if exec_options.reask_instructions:
-        instructions = Instructions(exec_options.reask_instructions)
-    if instructions is None:
-        instructions = Instructions("You are a helpful assistant.")
+    instructions = Instructions("You are a helpful assistant.")
     instructions = instructions.format(
         output_schema=schema_prompt_content,
         xml_output_schema=xml_output_schema,

guardrails/guard.py

@@ -841,6 +841,7 @@ def __call__(
         """
 
         messages = messages or self._exec_opts.messages or []
+        print("==== messages is", messages)
         # if messages is not None and not len(messages):
         #     raise RuntimeError(
         #         "You must provide messages. "

guardrails/llm_providers.py

@@ -416,16 +416,13 @@ def _invoke_llm(self, prompt: str, pipeline: Any, *args, **kwargs) -> LLMRespons
 class ArbitraryCallable(PromptCallableBase):
     def __init__(self, llm_api: Optional[Callable] = None, *args, **kwargs):
         llm_api_args = inspect.getfullargspec(llm_api)
+        if not llm_api_args.args:
+            raise ValueError(
+                "Custom LLM callables must accept"
+                " at least one positional argument for messages!"
+            )
         if not llm_api_args.varkw:
             raise ValueError("Custom LLM callables must accept **kwargs!")
-        if not llm_api_args.kwonlyargs or "messages" not in llm_api_args.kwonlyargs:
-            warnings.warn(
-                "We recommend including 'messages'"
-                " as keyword-only arguments for custom LLM callables."
-                " Doing so ensures these arguments are not unintentionally"
-                " passed through to other calls via **kwargs.",
-                UserWarning,
-            )
         self.llm_api = llm_api
         super().__init__(*args, **kwargs)
 
@@ -766,16 +763,13 @@ async def invoke_llm(
 class AsyncArbitraryCallable(AsyncPromptCallableBase):
     def __init__(self, llm_api: Callable, *args, **kwargs):
         llm_api_args = inspect.getfullargspec(llm_api)
+        if not llm_api_args.args:
+            raise ValueError(
+                "Custom LLM callables must accept"
+                " at least one positional argument for messages!"
+            )
         if not llm_api_args.varkw:
             raise ValueError("Custom LLM callables must accept **kwargs!")
-        if not llm_api_args.kwonlyargs or "messages" not in llm_api_args.kwonlyargs:
-            warnings.warn(
-                "We recommend including 'messages'"
-                " as keyword-only arguments for custom LLM callables."
-                " Doing so ensures these arguments are not uninentionally"
-                " passed through to other calls via **kwargs.",
-                UserWarning,
-            )
         self.llm_api = llm_api
         super().__init__(*args, **kwargs)
 

guardrails/prompt/messages.py

@@ -62,12 +62,16 @@ def format(
         """Format the messages using the given keyword arguments."""
         formatted_messages = []
         for message in self.source:
+            if isinstance(message["content"], str):
+                msg_str = message["content"]
+            else:
+                msg_str = message["content"]._source
             # Only use the keyword arguments that are present in the message.
-            vars = get_template_variables(message["content"])
+            vars = get_template_variables(msg_str)
             filtered_kwargs = {k: v for k, v in kwargs.items() if k in vars}
 
             # Return another instance of the class with the formatted message.
-            formatted_message = Template(message["content"]).safe_substitute(
+            formatted_message = Template(msg_str).safe_substitute(
                 **filtered_kwargs
             )
             formatted_messages.append(

guardrails/run/async_stream_runner.py

@@ -200,15 +200,16 @@ async def async_step(
     def get_chunk_text(self, chunk: Any, api: Union[PromptCallableBase, None]) -> str:
         """Get the text from a chunk."""
         chunk_text = ""
+    
         try:
             finished = chunk.choices[0].finish_reason
-            content = chunk.choices[0].text
+            content = chunk.choices[0].delta.content
             if not finished and content:
                 chunk_text = content
         except Exception:
             try:
                 finished = chunk.choices[0].finish_reason
-                content = chunk.choices[0].delta.content
+                content = chunk.choices[0].text
                 if not finished and content:
                     chunk_text = content
             except Exception:

guardrails/run/runner.py

@@ -35,7 +35,6 @@
 from guardrails.actions.reask import NonParseableReAsk, ReAsk, introspect
 from guardrails.telemetry import trace_call, trace_step
 
-
 class Runner:
     """Runner class that calls an LLM API with a prompt, and performs input and
     output validation.

guardrails/run/stream_runner.py

@@ -5,6 +5,7 @@
 from guardrails.classes.output_type import OT, OutputTypes
 from guardrails.classes.validation_outcome import ValidationOutcome
 from guardrails.llm_providers import (
+    LiteLLMCallable,
     PromptCallableBase,
 )
 from guardrails.run.runner import Runner
@@ -250,13 +251,13 @@ def get_chunk_text(self, chunk: Any, api: Union[PromptCallableBase, None]) -> st
         chunk_text = ""
         try:
             finished = chunk.choices[0].finish_reason
-            content = chunk.choices[0].text
+            content = chunk.choices[0].delta.content
             if not finished and content:
                 chunk_text = content
         except Exception:
             try:
                 finished = chunk.choices[0].finish_reason
-                content = chunk.choices[0].delta.content
+                content = chunk.choices[0].text
                 if not finished and content:
                     chunk_text = content
             except Exception:

tests/integration_tests/mock_llm_outputs.py

@@ -66,7 +66,7 @@ def _invoke_llm(
         self,
         prompt=None,
         instructions=None,
-        msg_history=None,
+        messages=None,
         base_model=None,
         *args,
         **kwargs,
@@ -128,28 +128,26 @@ def _invoke_llm(
         }
 
         try:
-            if msg_history:
-                key = (msg_history[0]["content"], msg_history[1]["content"])
-                print("=========trying key", key)
+            if messages:
+                key = (messages[0]["content"], messages[1]["content"])
                 out_text = mock_llm_responses[key]
-                print("========found out text", out_text)
-            if prompt and instructions and not msg_history:
+            if prompt and instructions and not messages:
                 out_text = mock_llm_responses[(prompt, instructions)]
-            elif msg_history and not prompt and not instructions:
-                if msg_history == entity_extraction.COMPILED_MSG_HISTORY:
+            elif messages and not prompt and not instructions:
+                if messages == entity_extraction.COMPILED_MSG_HISTORY:
                     out_text = entity_extraction.LLM_OUTPUT
                 elif (
-                    msg_history == string.MOVIE_MSG_HISTORY
+                    messages == string.MOVIE_MSG_HISTORY
                     and base_model == pydantic.WITH_MSG_HISTORY
                 ):
                     out_text = pydantic.MSG_HISTORY_LLM_OUTPUT_INCORRECT
-                elif msg_history == string.MOVIE_MSG_HISTORY:
+                elif messages == string.MOVIE_MSG_HISTORY:
                     out_text = string.MSG_LLM_OUTPUT_INCORRECT
                 else:
-                    raise ValueError("msg_history not found")
+                    raise ValueError("messages not found")
             else:
                 raise ValueError(
-                    "specify either prompt and instructions " "or msg_history"
+                    "specify either prompt and instructions " "or messages"
                 )
             return LLMResponse(
                 output=out_text,
@@ -160,7 +158,7 @@ def _invoke_llm(
             print("Unrecognized prompt!")
             print("\n prompt: \n", prompt)
             print("\n instructions: \n", instructions)
-            print("\n msg_history: \n", msg_history)
+            print("\n messages: \n", messages)
             raise ValueError("Compiled prompt not found")
 
 

tests/integration_tests/test_assets/custom_llm.py

@@ -2,16 +2,16 @@
 
 
 def mock_llm(
+    messages,
     *args,
-    messages: Optional[List[Dict[str, str]]] = None,
     **kwargs,
 ) -> str:
     return ""
 
 
 async def mock_async_llm(
+    messages,
     *args,
-    messages: Optional[List[Dict[str, str]]] = None,
     **kwargs,
 ) -> str:
     return ""

tests/integration_tests/test_assets/string/compiled_prompt.txt

@@ -2,7 +2,3 @@
 Given the following ingredients, what would you call this pizza?
 
 tomato, cheese, sour cream
-
-
-String Output:
-