Merge remote-tracking branch 'origin/main' into grace-sng7/main

Author: amit-sharma

pywhyllm/suggesters/identification_suggester.py

@@ -4,7 +4,7 @@
 import guidance
 from guidance import system, user, assistant, gen
 import re
-
+from inspect import cleandoc
 
 class IdentificationSuggester(IdentifierProtocol):
     CONTEXT: str = """causal mechanisms"""
@@ -199,14 +199,15 @@ def request_mediators(
             try:
                 lm = self.llm
                 with system():
-                    lm += f"""You are an expert in {domain_expertise} and are studying {analysis_context}. You are using your 
+                    prompt_str = f"""You are an expert in {domain_expertise} and are studying {analysis_context}. You are using your 
                 knowledge to help build a causal model that contains all the assumptions about the factors that are directly 
                 influencing athe {outcome}. Where a causal model is a conceptual model that describes the causal mechanisms 
                 of a system. You will do this by by answering questions about cause and effect and using your domain knowledge 
                 in {domain_expertise}. Follow the next two steps, and complete the first one before moving on to the second:"""
+                    lm += cleandoc(prompt_str)
 
                 with user():
-                    lm += f"""(1) From your perspective as an expert in {domain_expertise}, think step by step as you consider the factors 
+                    prompt_str = f"""(1) From your perspective as an expert in {domain_expertise}, think step by step as you consider the factors 
                 that may interact between the {treatment} and the {outcome}. Use your knowledge as an expert in 
                 {domain_expertise} to describe the mediators, if there are any at all, between {treatment} and the 
                 {outcome}. Be concise and keep your thinking within two paragraphs. Then provide your step by step chain 
@@ -223,6 +224,7 @@ def request_mediators(
                   the tags <mediating_factor>factor_name</mediating_factor>. Where factor_name is one of the items within the 
                   factor_names list. If a factor does not have a high likelihood of mediating, then do not wrap the factor with 
                   any tags. Your step by step answer as an in {domain_expertise}:"""
+                    lm += cleandoc(prompt_str)
 
                 with assistant():
                     lm += gen("output")
@@ -314,15 +316,16 @@ def request_ivs(
             try:
                 lm = self.llm
                 with system():
-                    lm += f"""You are an expert in {domain_expertise} and are studying {analysis_context}. 
+                    prompt_str = f"""You are an expert in {domain_expertise} and are studying {analysis_context}. 
                         You are using your knowledge to help build a causal model that contains all the assumptions about the factors 
                         that are directly influencing the {outcome}. Where a causal model is a conceptual model that describes the 
                         causal mechanisms of a system. You will do this by by answering questions about cause and effect and using 
                         your domain knowledge in {domain_expertise}. Follow the next two steps, and complete the first one before 
                         moving on to the second:"""
+                    lm += cleandoc(prompt_str)
 
                 with user():
-                    lm += f"""(1) From your perspective as an expert in {domain_expertise}, think step by step 
+                    prompt_str = f"""(1) From your perspective as an expert in {domain_expertise}, think step by step 
                        as you consider the factors that may interact with the {treatment} and do not interact with {outcome}. 
                        Use your knowlegde as an expert in {domain_expertise} to describe the instrumental variable(s), 
                        if there are any at all, that both has/have a high likelihood of influecing and causing the {treatment} and 
@@ -341,6 +344,7 @@ def request_ivs(
                        Where factor_name is one of the items within the factor_names list. If a factor does not have a high 
                        likelihood of being an instrumental variable, then do not wrap the factor with any tags. Your step by step 
                        answer as an in {domain_expertise}:"""
+                    lm += cleandoc(prompt_str)
                 with assistant():
                     lm += gen("output")
 

pywhyllm/suggesters/model_suggester.py

@@ -5,7 +5,7 @@
 from guidance import system, user, assistant, gen
 from ..helpers import RelationshipStrategy
 import re
-
+from inspect import cleandoc
 
 class ModelSuggester(ModelerProtocol):
     CONTEXT: str = """causal mechanisms"""
@@ -32,13 +32,14 @@ def suggest_domain_expertises(
                 with system():
                     lm += f"""You are a helpful assistant for recommending useful domain expertises."""
                 with user():
-                    lm += f"""What domain expertises have the knowledge and experience needed to identify causal 
+                    prompt_str = f"""What domain expertises have the knowledge and experience needed to identify causal 
                     relationships and causal influences between the {analysis_context}? What domain expertises are needed 
                     to work with and reason about the causal influence between {factors_list}? What domain expertises 
                     have the knowledge and experience to reason and answer questions about influence and cause between 
                     such factors? Think about this in a step by step manner and recommend {n_experts} expertises and 
                     provide each one wrapped within the tags, <domain_expertise></domain_expertise>, along with the 
                     reasoning and explanation wrapped between the tags <explanation></explanation>."""
+                    lm += cleandoc(prompt_str)
                 with assistant():
                     lm += gen("output")
 
@@ -74,13 +75,14 @@ def suggest_domain_experts(
                 with system():
                     lm += f"""You are a helpful assistant for recommending useful domain experts."""
                 with user():
-                    lm += f"""What domain experts have the knowledge and experience needed to identify causal relationships 
+                    prompt_str = f"""What domain experts have the knowledge and experience needed to identify causal relationships 
                     and causal influences between the {analysis_context}? What experts are needed to work with and 
                     reason about the causal influence between {factors_list}? What domain experts have the knowledge 
                     and experience to reason and answer questions about influence and cause between such factors? Think 
                     about this in a step by step manner and recommend {n_experts} domain experts and provide each one 
                     wrapped within the tags, <domain_expert></domain_expert>, along with the reasoning and explanation 
                     wrapped between the tags <explanation></explanation>."""
+                    lm += cleandoc(prompt_str)
                 with assistant():
                     lm += gen("output")
 
@@ -118,12 +120,13 @@ def suggest_stakeholders(
                     lm += "You are a helpful assistant for recommending useful primary and secondary stakeholders."
 
                 with user():
-                    lm += f"""What stakeholders have knowledge and experience in and about {analysis_context}? 
-                                    What stakeholders can work best with and reason well about the causal influence between 
+                    prompt_str = f"""What stakeholders have knowledge and experience in and about {analysis_context}? 
+                What stakeholders can work best with and reason well about the causal influence between 
                 {factors_list}? What stakeholders have the knowledge and experience useful to reason within this context? Think about 
                 this in a step by step manner and recommend {n_stakeholders} stakeholders. Then provide each useful stakeholder 
                 wrapped within the tags, <stakeholder></stakeholder>, along with the reasoning and explanation wrapped between the tags
                 <explanation></explanation>."""
+                    lm += cleandoc(prompt_str)
                 with assistant():
                     lm += gen("output")
 
@@ -204,13 +207,14 @@ def request_confounders(
             try:
                 lm = self.llm
                 with system():
-                    lm += f"""You are an expert in {domain_expertise} and are studying {analysis_context}.
-                You are using your knowledge to help build a causal model that contains all the assumptions about {
+                    prompt_str = f"""You are an expert in {domain_expertise} and are studying {analysis_context}.
+                    You are using your knowledge to help build a causal model that contains all the assumptions about {
                     analysis_context}. Where a causal model is a conceptual model that describes the causal mechanisms of a 
-                        system. You
-                will do this by by answering questions about cause and effect and using your domain knowledge in {domain_expertise}."""
+                    system. You
+                    will do this by answering questions about cause and effect and using your domain knowledge in {domain_expertise}."""
+                    lm += cleandoc(prompt_str)
                 with user():
-                    lm += f"""Follow the next two steps, and complete the first one before moving on to the second: (1) 
+                    prompt_str = f"""Follow the next two steps, and complete the first one before moving on to the second: (1) 
                                 From your perspective as an 
                 expert in {domain_expertise}, think step by step as you consider the factors that may interact between the {treatment} 
                 and the {outcome}. Use your knowlegde as an expert in {domain_expertise} to describe the confounders, if there are any 
@@ -230,6 +234,7 @@ def request_confounders(
                 <confounding_factor>factor_name</confounding_factor> where 
                 factor_name is one of the items within the factor_names list. If a factor does not have a high likelihood of directly 
                 confounding, then do not wrap the factor with any tags."""
+                    lm += cleandoc(prompt_str)
                 with assistant():
                     lm += gen("output")
 
@@ -284,7 +289,7 @@ def suggest_parents(
                     lm += f"""You are an expert in {domain_expertise} and are studying {analysis_context}"""
 
                 with user():
-                    lm += f"""You are using your knowledge to help build a causal model that 
+                    prompt_str = f"""You are using your knowledge to help build a causal model that 
                             contains all the assumptions about the factors that are directly influencing 
                             and causing the {factor}. Where a causal model is a conceptual model that describes the 
                             causal mechanisms of a system. You will do this by by answering questions about cause and 
@@ -306,6 +311,7 @@ def suggest_parents(
 {factor}, 
                             then do not wrap the factor with any tags. Your answer as an expert in 
 {domain_expertise}:"""
+                    lm += cleandoc(prompt_str)
 
                 with assistant():
                     lm += gen("output")
@@ -349,7 +355,7 @@ def suggest_children(
                 with system():
                     lm += f"""You are an expert in {domain_expertise} and are studying {analysis_context}"""
                 with user():
-                    lm += f"""You are using your knowledge to help build a causal model that 
+                    prompt_str = f"""You are using your knowledge to help build a causal model that 
                             contains all the assumptions about the factors that are directly influencing and causing the {factor}. 
                             Where a 
                             causal model is a conceptual model that describes the causal mechanisms of a system. You will do this by by 
@@ -377,7 +383,7 @@ def suggest_children(
                             factor_names list. If a factor does not have a high likelihood of directly influencing and causing the {
                     factor}, then do not wrap the factor with any tags. Your answer as an expert in 
                 {domain_expertise}:"""
-
+                    lm += cleandoc(prompt_str)
                 with assistant():
                     lm += gen("output")
 
@@ -410,16 +416,17 @@ def suggest_pairwise_relationship(
             try:
                 lm = self.llm
                 with system():
-                    lm += f"""You are an expert in {domain_expertise} and are 
+                    prompt_str = f"""You are an expert in {domain_expertise} and are 
                             studying {analysis_context}. You are using your knowledge to help build a causal model that contains 
                             all the 
                             assumptions about {analysis_context}. Where a causal model is a conceptual model that describes the 
                             causal 
                             mechanisms of a system. You will do this by by answering questions about cause and effect and using your 
                             domain 
                             knowledge as an expert in {domain_expertise}."""
+                    lm += cleandoc(prompt_str)
                 with user():
-                    lm += f"""From your perspective as an expert in {domain_expertise}, which of the following is 
+                    prompt_str = f"""From your perspective as an expert in {domain_expertise}, which of the following is 
                                     most likely true? (A) {factor_a} affects {factor_b}; {factor_a} has a high likelihood of directly 
                                     influencing {factor_b}; and {factor_a} causes {factor_b}. (B) {factor_b} affects {factor_a}; 
                 {factor_b} has a high likelihood of directly influencing {factor_a}; and {factor_b} causes {factor_a}. (C) Neither A 
@@ -429,6 +436,7 @@ def suggest_pairwise_relationship(
                 you reach a conclusion, wrap your answer within the tags <answer></answer>. If you are done thinking, provide your 
                 answer wrapped within the tags <answer></answer>. e.g. <answer>A</answer>, <answer>B</answer>, or <answer>C</answer>. 
                 Your answer as an expert in {domain_expertise}:"""
+                    lm += cleandoc(prompt_str)
 
                 with assistant():
                     lm += gen("output")

pywhyllm/suggesters/simple_identification_suggester.py

@@ -1,7 +1,7 @@
 import guidance
 from guidance import system, user, assistant, gen
 import re
-
+from inspect import cleandoc
 
 class SimpleIdentificationSuggester:
 
@@ -16,9 +16,10 @@ def suggest_iv(self, factors, treatment, outcome):
             lm += "You are a helpful assistant for causal reasoning."
 
         with user():
-            lm += f"""Which factors in {factors} might be valid instrumental variables for identifying the effect of {treatment} on {outcome}?
+            prompt_str = f"""Which factors in {factors} might be valid instrumental variables for identifying the effect of {treatment} on {outcome}?
 
             List the factors that are possible instrumental variables in <iv> </iv> tags."""
+            lm += cleandoc(prompt_str)
         with assistant():
             lm += gen("iv")
 
@@ -33,10 +34,11 @@ def suggest_backdoor(self, factors, treatment, outcome):
             lm += "You are a helpful assistant for causal reasoning."
 
         with user():
-            lm += f"""Which set or subset of factors in {factors} might satisfy the backdoor criteria for identifying the effect of {treatment} on {outcome}?
+            prompt_str = f"""Which set or subset of factors in {factors} might satisfy the backdoor criteria for identifying the effect of {treatment} on {outcome}?
 
             List the factors satisfying the backdoor criteria enclosing the name of each factor in <backdoor> </backdoor> tags.
             """
+            lm += cleandoc(prompt_str)
         with assistant():
             lm += gen("backdoors")
 
@@ -51,10 +53,11 @@ def suggest_frontdoor(self, factors, treatment, outcome):
             lm += "You are a helpful assistant for causal reasoning."
 
         with user():
-            lm += f"""Which set or subset of factors in {factors} might satisfy the frontdoor criteria for identifying the effect of {treatment} on {outcome}?
+            prompt_str = f"""Which set or subset of factors in {factors} might satisfy the frontdoor criteria for identifying the effect of {treatment} on {outcome}?
 
             List the factors satisfying the frontdoor criteria enclosing the name of each factor in <frontdoor> </frontdoor> tags.
             """
+            lm += cleandoc(prompt_str)
         with assistant():
             lm += gen("frontdoors")
 

pywhyllm/suggesters/simple_model_suggester.py

@@ -3,7 +3,7 @@
 import re
 import itertools
 from guidance import system, user, assistant, gen
-
+from inspect import cleandoc
 
 class SimpleModelSuggester:
     """
@@ -43,8 +43,9 @@ def suggest_pairwise_relationship(self, variable1: str, variable2: str):
             lm += "You are a helpful assistant for causal reasoning."
 
         with user():
-            lm += f"""Which cause-and-effect-relationship is more likely? Provide reasoning and give your final answer (A, B, or C) in <answer> </answer> tags with the letter only and no whitespaces.
+            prompt_str = f"""Which cause-and-effect-relationship is more likely? Provide reasoning and give your final answer (A, B, or C) in <answer> </answer> tags with the letter only and no whitespaces.
             A. {variable1} causes {variable2} B. {variable2} causes {variable1} C. neither {variable1} nor {variable2} cause each other."""
+            lm += cleandoc(prompt_str)
 
         with assistant():
             lm += gen("description")
@@ -110,12 +111,13 @@ def suggest_confounders(self, variables: List[str], treatment: str, outcome: str
             lm += "You are a helpful assistant for causal reasoning."
 
         with user():
-            lm += f"""What latent confounding factors might influence the relationship between {treatment} and {outcome}?
+            prompt_str = f"""What latent confounding factors might influence the relationship between {treatment} and {outcome}?
 
             We have already considered the following factors {variables}.  Please do not repeat them.
 
             List the confounding factors between {treatment} and {outcome} enclosing the name of each factor in <conf> </conf> tags.
             """
+            lm += cleandoc(prompt_str)
         with assistant():
             lm += gen("latents")