Multiple Selectable Approaches for Text (Chain of Thought with Prompt…

… Chaining) (#350)

Co-authored-by: = Enea_Gore <[email protected]>
Co-authored-by: Maximilian Sölch <[email protected]>

llm_core/llm_core/models/__init__.py

@@ -6,6 +6,7 @@
 
 
 DefaultModelConfig: Type[ModelConfig]
+MiniModelConfig: ModelConfig
 default_model_name = os.environ.get("LLM_DEFAULT_MODEL")
 evaluation_model_name = os.environ.get("LLM_EVALUATION_MODEL")
 
@@ -18,6 +19,8 @@
     types.append(openai_config.OpenAIModelConfig)
     if default_model_name in openai_config.available_models:
         DefaultModelConfig = openai_config.OpenAIModelConfig
+    if "openai_gpt-4o-mini" in openai_config.available_models:        
+        MiniModelConfig = openai_config.OpenAIModelConfig(model_name="openai_gpt-4o-mini",max_tokens=3000, temperature=0,top_p=0.9,presence_penalty=0,frequency_penalty=0)
     if evaluation_model_name in openai_config.available_models:
         evaluation_model = openai_config.available_models[evaluation_model_name]
 except AttributeError:

llm_core/llm_core/utils/llm_utils.py

@@ -1,7 +1,7 @@
 from typing import Type, TypeVar, List
 from pydantic import BaseModel
 import tiktoken
-from langchain.chat_models import ChatOpenAI
+from langchain_openai import AzureChatOpenAI, ChatOpenAI
 from langchain.base_language import BaseLanguageModel
 from langchain.prompts import (
     ChatPromptTemplate,
@@ -75,7 +75,7 @@ def supports_function_calling(model: BaseLanguageModel):
     Returns:
         boolean: True if the model supports function calling, False otherwise
     """
-    return isinstance(model, ChatOpenAI)
+    return isinstance(model, ChatOpenAI)  or isinstance(model, AzureChatOpenAI)
 
 
 def get_chat_prompt_with_formatting_instructions(

llm_core/llm_core/utils/predict_and_parse.py

@@ -12,7 +12,8 @@ async def predict_and_parse(
         chat_prompt: ChatPromptTemplate, 
         prompt_input: dict, 
         pydantic_object: Type[T], 
-        tags: Optional[List[str]]
+        tags: Optional[List[str]],
+        use_function_calling: bool = False
     ) -> Optional[T]:
     """Predicts an LLM completion using the model and parses the output using the provided Pydantic model
 
@@ -36,13 +37,30 @@ async def predict_and_parse(
     if experiment.run_id is not None:
         tags.append(f"run-{experiment.run_id}")
 
-    structured_output_llm = model.with_structured_output(pydantic_object, method="json_mode")
-    chain = RunnableSequence(
-        chat_prompt,
-        structured_output_llm
-    )
-
-    try:
-        return await chain.ainvoke(prompt_input, config={"tags": tags})
-    except ValidationError as e:
-        raise ValueError(f"Could not parse output: {e}") from e
+
+    if (use_function_calling):
+        structured_output_llm = model.with_structured_output(pydantic_object)
+        chain = chat_prompt | structured_output_llm
+
+        try:
+            result = await chain.ainvoke(prompt_input, config={"tags": tags})
+
+            if isinstance(result, pydantic_object):
+                return result
+            else:
+                raise ValueError("Parsed output does not match the expected Pydantic model.")
+
+        except ValidationError as e:
+            raise ValueError(f"Could not parse output: {e}") from e
+
+    else:
+        structured_output_llm = model.with_structured_output(pydantic_object, method = "json_mode")
+        chain = RunnableSequence(
+            chat_prompt,
+            structured_output_llm
+        )
+        try:
+            return await chain.ainvoke(prompt_input, config={"tags": tags})
+        except ValidationError as e:
+            raise ValueError(f"Could not parse output: {e}") from e
+

modules/text/module_text_llm/module_text_llm/__main__.py

@@ -1,19 +1,16 @@
-import json
 import os
 from typing import List, Any
 
 import nltk
 import tiktoken
-
 from athena import app, submission_selector, submissions_consumer, feedback_consumer, feedback_provider, evaluation_provider
 from athena.text import Exercise, Submission, Feedback
 from athena.logger import logger
 
 from module_text_llm.config import Configuration
 from module_text_llm.evaluation import get_feedback_statistics, get_llm_statistics
-from module_text_llm.generate_suggestions import generate_suggestions
 from module_text_llm.generate_evaluation import generate_evaluation
-
+from module_text_llm.approach_controller import generate_suggestions
 
 @submissions_consumer
 def receive_submissions(exercise: Exercise, submissions: List[Submission]):
@@ -30,7 +27,6 @@ def select_submission(exercise: Exercise, submissions: List[Submission]) -> Subm
 def process_incoming_feedback(exercise: Exercise, submission: Submission, feedbacks: List[Feedback]):
     logger.info("process_feedback: Received %d feedbacks for submission %d of exercise %d.", len(feedbacks), submission.id, exercise.id)
 
-
 @feedback_provider
 async def suggest_feedback(exercise: Exercise, submission: Submission, is_graded: bool, module_config: Configuration) -> List[Feedback]:
     logger.info("suggest_feedback: %s suggestions for submission %d of exercise %d were requested",

modules/text/module_text_llm/module_text_llm/approach_config.py

@@ -0,0 +1,16 @@
+from abc import ABC
+from pydantic import BaseModel, Field
+from llm_core.models import ModelConfigType, DefaultModelConfig
+from enum import Enum
+
+class ApproachType(str, Enum):
+    basic = "BasicApproach"
+    chain_of_thought = "ChainOfThought"
+
+class ApproachConfig(BaseModel, ABC):
+    max_input_tokens: int = Field(default=3000, description="Maximum number of tokens in the input prompt.")
+    model: ModelConfigType = Field(default=DefaultModelConfig())
+    type: str = Field(..., description="The type of approach config")
+
+    class Config:
+        use_enum_values = True

modules/text/module_text_llm/module_text_llm/approach_controller.py

@@ -0,0 +1,16 @@
+
+from typing import List
+from athena.text import Exercise, Submission, Feedback
+from module_text_llm.basic_approach import BasicApproachConfig
+from module_text_llm.chain_of_thought_approach  import ChainOfThoughtConfig
+from module_text_llm.approach_config import ApproachConfig
+
+from module_text_llm.basic_approach.generate_suggestions import generate_suggestions as generate_suggestions_basic
+from module_text_llm.chain_of_thought_approach.generate_suggestions import generate_suggestions as generate_cot_suggestions
+
+async def generate_suggestions(exercise: Exercise, submission: Submission, config: ApproachConfig, debug: bool) -> List[Feedback]:
+    if(isinstance(config, BasicApproachConfig)):
+        return await generate_suggestions_basic(exercise, submission, config, debug)
+    elif(isinstance(config, ChainOfThoughtConfig)):
+        return await generate_cot_suggestions(exercise, submission, config, debug)
+

modules/text/module_text_llm/module_text_llm/basic_approach/__init__.py

@@ -0,0 +1,11 @@
+from module_text_llm.approach_config import ApproachConfig
+from pydantic import Field
+from typing import Literal
+
+
+from module_text_llm.basic_approach.prompt_generate_suggestions import GenerateSuggestionsPrompt
+
+class BasicApproachConfig(ApproachConfig):
+    type: Literal['basic'] = 'basic'
+    generate_suggestions_prompt: GenerateSuggestionsPrompt = Field(default=GenerateSuggestionsPrompt())
+

modules/text/module_text_llm/module_text_llm/generate_suggestions.py → modules/text/module_text_llm/module_text_llm/basic_approach/generate_suggestions.py

@@ -1,46 +1,22 @@
-from typing import List, Optional, Sequence
-from pydantic import BaseModel, Field
+from typing import List
 
 from athena import emit_meta
 from athena.text import Exercise, Submission, Feedback
 from athena.logger import logger
-
-from module_text_llm.config import BasicApproachConfig
 from llm_core.utils.llm_utils import (
     get_chat_prompt_with_formatting_instructions, 
     check_prompt_length_and_omit_features_if_necessary, 
     num_tokens_from_prompt,
 )
+from athena.text import Exercise, Submission, Feedback
 from llm_core.utils.predict_and_parse import predict_and_parse
 
+from module_text_llm.config import BasicApproachConfig
 from module_text_llm.helpers.utils import add_sentence_numbers, get_index_range_from_line_range, format_grading_instructions
-
-class FeedbackModel(BaseModel):
-    title: str = Field(description="Very short title, i.e. feedback category or similar", example="Logic Error")
-    description: str = Field(description="Feedback description")
-    line_start: Optional[int] = Field(description="Referenced line number start, or empty if unreferenced")
-    line_end: Optional[int] = Field(description="Referenced line number end, or empty if unreferenced")
-    credits: float = Field(0.0, description="Number of points received/deducted")
-    grading_instruction_id: Optional[int] = Field(
-        description="ID of the grading instruction that was used to generate this feedback, or empty if no grading instruction was used"
-    )
-
-    class Config:
-        title = "Feedback"
-
-
-class AssessmentModel(BaseModel):
-    """Collection of feedbacks making up an assessment"""
-
-    feedbacks: Sequence[FeedbackModel] = Field(description="Assessment feedbacks")
-
-    class Config:
-        title = "Assessment"
-
+from module_text_llm.basic_approach.prompt_generate_suggestions import AssessmentModel
 
 async def generate_suggestions(exercise: Exercise, submission: Submission, config: BasicApproachConfig, debug: bool) -> List[Feedback]:
     model = config.model.get_model()  # type: ignore[attr-defined]
-
     prompt_input = {
         "max_points": exercise.max_points,
         "bonus_points": exercise.bonus_points,
@@ -83,7 +59,8 @@ async def generate_suggestions(exercise: Exercise, submission: Submission, confi
         tags=[
             f"exercise-{exercise.id}",
             f"submission-{submission.id}",
-        ]
+        ],
+        use_function_calling=True
     )
 
     if debug:

modules/text/module_text_llm/module_text_llm/basic_approach/prompt_generate_suggestions.py

@@ -0,0 +1,65 @@
+from pydantic import Field, BaseModel
+from typing import List, Optional
+from pydantic import BaseModel, Field
+
+system_message = """\
+You are an AI tutor for text assessment at a prestigious university.
+
+# Task
+Create graded feedback suggestions for a student\'s text submission that a human tutor would accept. \
+Meaning, the feedback you provide should be applicable to the submission with little to no modification.
+
+# Style
+1. Constructive, 2. Specific, 3. Balanced, 4. Clear and Concise, 5. Actionable, 6. Educational, 7. Contextual
+
+# Problem statement
+{problem_statement}
+
+# Example solution
+{example_solution}
+
+# Grading instructions
+{grading_instructions}
+Max points: {max_points}, bonus points: {bonus_points}\
+    
+Respond in json.
+"""
+
+human_message = """\
+Student\'s submission to grade (with sentence numbers <number>: <sentence>):
+
+Respond in json.
+
+\"\"\"
+{submission}
+\"\"\"\
+"""
+
+# Input Prompt
+class GenerateSuggestionsPrompt(BaseModel):
+    """\
+Features available: **{problem_statement}**, **{example_solution}**, **{grading_instructions}**, **{max_points}**, **{bonus_points}**, **{submission}**
+
+_Note: **{problem_statement}**, **{example_solution}**, or **{grading_instructions}** might be omitted if the input is too long._\
+"""
+    system_message: str = Field(default=system_message,
+                                description="Message for priming AI behavior and instructing it what to do.")
+    human_message: str = Field(default=human_message,
+                               description="Message from a human. The input on which the AI is supposed to act.")
+# Output Object
+class FeedbackModel(BaseModel):
+    title: str = Field(description="Very short title, i.e. feedback category or similar", example="Logic Error")
+    description: str = Field(description="Feedback description")
+    line_start: Optional[int] = Field(description="Referenced line number start, or empty if unreferenced")
+    line_end: Optional[int] = Field(description="Referenced line number end, or empty if unreferenced")
+    credits: float = Field(0.0, description="Number of points received/deducted")
+    grading_instruction_id: Optional[int] = Field(
+        description="ID of the grading instruction that was used to generate this feedback, or empty if no grading instruction was used"
+    )
+
+
+class AssessmentModel(BaseModel):
+    """Collection of feedbacks making up an assessment"""
+
+    feedbacks: List[FeedbackModel] = Field(description="Assessment feedbacks")
+

modules/text/module_text_llm/module_text_llm/chain_of_thought_approach/__init__.py

@@ -0,0 +1,15 @@
+from pydantic import BaseModel, Field
+from typing import Literal
+from llm_core.models import ModelConfigType, MiniModelConfig
+
+from module_text_llm.approach_config import ApproachConfig
+from module_text_llm.chain_of_thought_approach.prompt_generate_feedback import CoTGenerateSuggestionsPrompt
+from module_text_llm.chain_of_thought_approach.prompt_thinking import ThinkingPrompt
+
+class ChainOfThoughtConfig(ApproachConfig):
+    # Defaults to the cheaper mini 4o model
+    type: Literal['chain_of_thought'] = 'chain_of_thought'
+    model: ModelConfigType = Field(default=MiniModelConfig)  # type: ignore
+    thikning_prompt: ThinkingPrompt = Field(default=ThinkingPrompt())
+    generate_suggestions_prompt: CoTGenerateSuggestionsPrompt = Field(default=CoTGenerateSuggestionsPrompt())
+