@@ -30,6 +30,24 @@ def filter(self, instruction: str, model):
3030 """
3131 Filters the DataFrame with the semantics of the user instruction.
3232
33+ **Examples:**
34+
35+ >>> import bigframes.pandas as bpd
36+ >>> bpd.options.display.progress_bar = None
37+
38+ >>> import bigframes
39+ >>> bigframes.options.experiments.semantic_operators = True
40+
41+ >>> import bigframes.ml.llm as llm
42+ >>> model = llm.GeminiTextGenerator(model_name="gemini-1.5-flash-001")
43+
44+ >>> df = bpd.DataFrame({"country": ["USA", "Germany"], "city": ["Seattle", "Berlin"]})
45+ >>> df.semantics.filter("{city} is the capital of {country}", model)
46+ country city
47+ 1 Germany Berlin
48+ <BLANKLINE>
49+ [1 rows x 2 columns]
50+
3351 Args:
3452 instruction:
3553 An instruction on how to filter the data. This value must contain
@@ -39,7 +57,7 @@ def filter(self, instruction: str, model):
3957 "The {food} is healthy."
4058
4159 model:
42- A LLM model provided by Bigframes ML package.
60+ A GeminiTextGenerator provided by Bigframes ML package.
4361
4462 Returns:
4563 DataFrame filtered by the instruction.
@@ -49,9 +67,89 @@ def filter(self, instruction: str, model):
4967 ValueError: when the instruction refers to a non-existing column, or when no
5068 columns are referred to.
5169 """
70+ _validate_model (model )
71+
72+ output_instruction = "Based on the provided context, reply to the following claim by only True or False:"
73+
74+ from bigframes .dataframe import DataFrame
75+
76+ results = typing .cast (
77+ DataFrame , model .predict (self ._make_prompt (instruction , output_instruction ))
78+ )
79+
80+ return self ._df [
81+ results ["ml_generate_text_llm_result" ].str .lower ().str .contains ("true" )
82+ ]
83+
84+ def map (self , instruction : str , output_column : str , model ):
85+ """
86+ Maps the DataFrame with the semantics of the user instruction.
87+
88+ **Examples:**
89+
90+ >>> import bigframes.pandas as bpd
91+ >>> bpd.options.display.progress_bar = None
92+
93+ >>> import bigframes
94+ >>> bigframes.options.experiments.semantic_operators = True
95+
96+ >>> import bigframes.ml.llm as llm
97+ >>> model = llm.GeminiTextGenerator(model_name="gemini-1.5-flash-001")
98+
99+ >>> df = bpd.DataFrame({"ingredient_1": ["Burger Bun", "Soy Bean"], "ingredient_2": ["Beef Patty", "Bittern"]})
100+ >>> df.semantics.map("What is the food made from {ingredient_1} and {ingredient_2}? One word only.", result_column_name="food", model=model)
101+ ingredient_1 ingredient_2 food
102+ 0 Burger Bun Beef Patty Burger
103+ <BLANKLINE>
104+ 1 Soy Bean Bittern Tofu
105+ <BLANKLINE>
106+ <BLANKLINE>
107+ [2 rows x 3 columns]
108+
109+ Args:
110+ instruction:
111+ An instruction on how to map the data. This value must contain
112+ column references by name, which should be wrapped in a pair of braces.
113+ For example, if you have a column "food", you can refer to this column
114+ in the instructions like:
115+ "Get the ingredients of {food}."
52116
117+ result_column_name:
118+ The column name of the mapping result.
119+
120+ model:
121+ A GeminiTextGenerator provided by Bigframes ML package.
122+
123+ Returns:
124+ DataFrame with attached mapping results.
125+
126+ Raises:
127+ NotImplementedError: when the semantic operator experiment is off.
128+ ValueError: when the instruction refers to a non-existing column, or when no
129+ columns are referred to.
130+ """
131+ _validate_model (model )
132+
133+ output_instruction = (
134+ "Based on the provided contenxt, answer the following instruction:"
135+ )
136+
137+ from bigframes .series import Series
138+
139+ results = typing .cast (
140+ Series ,
141+ model .predict (self ._make_prompt (instruction , output_instruction ))[
142+ "ml_generate_text_llm_result"
143+ ],
144+ )
145+
146+ from bigframes .core .reshape import concat
147+
148+ return concat ([self ._df , results .rename (output_column )], axis = 1 )
149+
150+ def _make_prompt (self , user_instruction : str , output_instruction : str ):
53151 # Validate column references
54- columns = re .findall (r"(?<!{)\{(?!{)(.*?)\}(?!\})" , instruction )
152+ columns = re .findall (r"(?<!{)\{(?!{)(.*?)\}(?!\})" , user_instruction )
55153
56154 if not columns :
57155 raise ValueError ("No column references." )
@@ -61,30 +159,20 @@ def filter(self, instruction: str, model):
61159 raise ValueError (f"Column { column } )
62160
63161 # Replace column references with names.
64- instruction = instruction .format (** {col : col for col in columns })
162+ user_instruction = user_instruction .format (** {col : col for col in columns })
65163
66- prompt_df = self ._df .copy ()
164+ prompt_df = self ._df [columns ].copy ()
165+ prompt_df ["prompt" ] = f"{ output_instruction } \n { user_instruction } \n Context: "
67166
68167 # Combine context from multiple columns.
69- for idx , col in enumerate (columns ):
70- if idx == 0 :
71- prompt_df ["context" ] = f"{ col } + prompt_df [col ] + "`\n "
72- else :
73- prompt_df ["context" ] += f"{ col } + prompt_df [col ] + "`\n "
74-
75- prompt_df ["prompt" ] = (
76- "Decide the folowing claim by only True and False: "
77- + instruction
78- + "\n Context:"
79- + prompt_df ["context" ]
80- )
168+ for col in columns :
169+ prompt_df ["prompt" ] += f"{ col } + prompt_df [col ] + "`\n "
81170
82- import bigframes . dataframe
171+ return prompt_df [ "prompt" ]
83172
84- results = typing .cast (
85- bigframes .dataframe .DataFrame , model .predict (prompt_df ["prompt" ])
86- )
87173
88- return self ._df [
89- results ["ml_generate_text_llm_result" ].str .lower ().str .contains ("true" )
90- ]
174+ def _validate_model (model ):
175+ from bigframes .ml .llm import GeminiTextGenerator
176+
177+ if not isinstance (model , GeminiTextGenerator ):
178+ raise ValueError ("Model is not GeminiText Generator" )
![gcf-owl-bot[bot]](https://avatars.githubusercontent.com/in/99011?v=4&size=40){kind=avatar}
0 commit comments