adjusted prompt to reduce halucinations: nested filter/context calls and putting filter args in quotation marks

Author: ds-jakub-cierocki

src/dbally/iql_generator/prompt.py

@@ -76,10 +76,16 @@ def __init__(
                 "It is VERY IMPORTANT not to use methods other than those listed above.\n"
                 "Finally, if a called function argument value is not directly specified in the query but instead requires "
                 "some additional execution context, than substitute that argument value with: AskerContext().\n"
-                'The typical input phrase suggesting that the additional execution context need to be referenced contains words like: "I", "my", "mine", "current", "the" etc..\n'
+                "The typical input phrase suggesting that the additional execution context need to be referenced \n"
+                'contains words like: "I", "my", "mine", "current", "the" etc..\n'
                 'For example: "my position name", "my company valuation", "current day", "the ongoing project".\n'
                 "In that case, the part of the output will look like this:\n"
                 "filter4(AskerContext())\n"
+                "Outside this situation DO NOT combine filters like this:\n"
+                "filter4(filter2())\n"
+                "And NEVER quote the filter argument unless you're sure it represents the string/literal datatype, \n"
+                "Especially do not quote AskerContext() calls like this:\n"
+                "filter2('AskerContext()')\n"
                 """If you DON'T KNOW HOW TO ANSWER DON'T SAY \"\", SAY: `UNSUPPORTED QUERY` INSTEAD! """
                 "This is CRUCIAL, otherwise the system will crash. "
             ),

tests/unit/test_iql_format.py

@@ -22,12 +22,19 @@ async def test_iql_prompt_format_default() -> None:
             "DO NOT INCLUDE arguments names in your response. Only the values.\n"
             "You MUST use only these methods:\n"
             "\n\n"
-            "It is VERY IMPORTANT not to use methods other than those listed above."
-            "Finally, if a called function argument value is not directly specified in the query but instead requires some additional execution context, than substitute that argument value with: AskerContext()."
-            'The typical input phrase suggesting that the additional execution context need to be referenced contains words like: "I", "my", "mine", "current", "the" etc..'
-            'For example: "my position name", "my company valuation", "current day", "the ongoing project".'
-            "In that case, the part of the output will look like this:"
-            "filter4(AskerContext())"
+            "It is VERY IMPORTANT not to use methods other than those listed above.\n"
+            "Finally, if a called function argument value is not directly specified in the query but instead requires "
+            "some additional execution context, than substitute that argument value with: AskerContext().\n"
+            "The typical input phrase suggesting that the additional execution context need to be referenced \n"
+            'contains words like: "I", "my", "mine", "current", "the" etc..\n'
+            'For example: "my position name", "my company valuation", "current day", "the ongoing project".\n'
+            "In that case, the part of the output will look like this:\n"
+            "filter4(AskerContext())\n"
+            "Outside this situation DO NOT combine filters like this:\n"
+            "filter4(filter2())\n"
+            "And NEVER quote the filter argument unless you're sure it represents the string/literal datatype, \n"
+            "Especially do not quote AskerContext() calls like this:\n"
+            "filter2('AskerContext()')\n"
             """If you DON'T KNOW HOW TO ANSWER DON'T SAY \"\", SAY: `UNSUPPORTED QUERY` INSTEAD! """
             "This is CRUCIAL, otherwise the system will crash. ",
             "is_example": False,
@@ -57,12 +64,19 @@ async def test_iql_prompt_format_few_shots_injected() -> None:
             "DO NOT INCLUDE arguments names in your response. Only the values.\n"
             "You MUST use only these methods:\n"
             "\n\n"
-            "It is VERY IMPORTANT not to use methods other than those listed above."
-            "Finally, if a called function argument value is not directly specified in the query but instead requires some additional execution context, than substitute that argument value with: AskerContext()."
-            'The typical input phrase suggesting that the additional execution context need to be referenced contains words like: "I", "my", "mine", "current", "the" etc..'
-            'For example: "my position name", "my company valuation", "current day", "the ongoing project".'
-            "In that case, the part of the output will look like this:"
-            "filter4(AskerContext())"
+            "It is VERY IMPORTANT not to use methods other than those listed above.\n"
+            "Finally, if a called function argument value is not directly specified in the query but instead requires "
+            "some additional execution context, than substitute that argument value with: AskerContext().\n"
+            "The typical input phrase suggesting that the additional execution context need to be referenced \n"
+            'contains words like: "I", "my", "mine", "current", "the" etc..\n'
+            'For example: "my position name", "my company valuation", "current day", "the ongoing project".\n'
+            "In that case, the part of the output will look like this:\n"
+            "filter4(AskerContext())\n"
+            "Outside this situation DO NOT combine filters like this:\n"
+            "filter4(filter2())\n"
+            "And NEVER quote the filter argument unless you're sure it represents the string/literal datatype, \n"
+            "Especially do not quote AskerContext() calls like this:\n"
+            "filter2('AskerContext()')\n"
             """If you DON'T KNOW HOW TO ANSWER DON'T SAY \"\", SAY: `UNSUPPORTED QUERY` INSTEAD! """
             "This is CRUCIAL, otherwise the system will crash. ",
             "is_example": False,