@@ -41,13 +41,21 @@ class ChatReadRetrieveReadApproach:
4141If you cannot generate a search query, return just the number 0.
4242"""
4343 query_prompt_few_shots = [
44- {' role' : USER , ' content' : ' What are my health plans?' },
45- {' role' : ASSISTANT , ' content' : ' Show available health plans' },
46- {' role' : USER , ' content' : ' does my plan cover cardio?' },
47- {' role' : ASSISTANT , ' content' : ' Health plan cardio coverage' }
44+ {" role" USER , " content" : " What are my health plans?"
45+ {" role" ASSISTANT , " content" : " Show available health plans"
46+ {" role" USER , " content" : " does my plan cover cardio?"
47+ {" role" ASSISTANT , " content" : " Health plan cardio coverage" },
4848 ]
4949
50- def __init__ (self , search_client : SearchClient , chatgpt_deployment : str , chatgpt_model : str , embedding_deployment : str , sourcepage_field : str , content_field : str ):
50+ def __init__ (
51+ self ,
52+ search_client : SearchClient ,
53+ chatgpt_deployment : str ,
54+ chatgpt_model : str ,
55+ embedding_deployment : str ,
56+ sourcepage_field : str ,
57+ content_field : str ,
58+ ):
5159 self .search_client = search_client
5260 self .chatgpt_deployment = chatgpt_deployment
5361 self .chatgpt_model = chatgpt_model
@@ -56,15 +64,17 @@ def __init__(self, search_client: SearchClient, chatgpt_deployment: str, chatgpt
5664 self .content_field = content_field
5765 self .chatgpt_token_limit = get_token_limit (chatgpt_model )
5866
59- async def run_until_final_call (self , history : list [dict [str , str ]], overrides : dict [str , Any ], should_stream : bool = False ) -> tuple :
67+ async def run_until_final_call (
68+ self , history : list [dict [str , str ]], overrides : dict [str , Any ], should_stream : bool = False
69+ ) -> tuple :
6070 has_text = overrides .get ("retrieval_mode" ) in ["text" , "hybrid" , None ]
6171 has_vector = overrides .get ("retrieval_mode" ) in ["vectors" , "hybrid" , None ]
6272 use_semantic_captions = True if overrides .get ("semantic_captions" ) and has_text else False
6373 top = overrides .get ("top" ) or 3
6474 exclude_category = overrides .get ("exclude_category" ) or None
6575 filter = "category ne '{}'" .format (exclude_category .replace ("'" , "''" )) if exclude_category else None
6676
67- user_q = ' Generate search query for: ' + history [- 1 ]["user" ]
77+ user_q = " Generate search query for: " + history [- 1 ]["user" ]
6878
6979 # STEP 1: Generate an optimized keyword search query based on the chat history and the last question
7080 messages = self .get_messages_from_history (
@@ -73,89 +83,112 @@ async def run_until_final_call(self, history: list[dict[str, str]], overrides: d
7383 history ,
7484 user_q ,
7585 self .query_prompt_few_shots ,
76- self .chatgpt_token_limit - len (user_q )
77- )
86+ self .chatgpt_token_limit - len (user_q ),
87+ )
7888
7989 chat_completion = await openai .ChatCompletion .acreate (
8090 deployment_id = self .chatgpt_deployment ,
8191 model = self .chatgpt_model ,
8292 messages = messages ,
8393 temperature = 0.0 ,
8494 max_tokens = 32 ,
85- n = 1 )
95+ n = 1 ,
96+ )
8697
8798 query_text = chat_completion .choices [0 ].message .content
8899 if query_text .strip () == "0" :
89- query_text = history [- 1 ]["user" ] # Use the last user input if we failed to generate a better query
100+ query_text = history [- 1 ]["user" ] # Use the last user input if we failed to generate a better query
90101
91102 # STEP 2: Retrieve relevant documents from the search index with the GPT optimized query
92103
93104 # If retrieval mode includes vectors, compute an embedding for the query
94105 if has_vector :
95- query_vector = (await openai .Embedding .acreate (engine = self .embedding_deployment , input = query_text ))["data" ][0 ]["embedding" ]
106+ embedding = await openai .Embedding .acreate (engine = self .embedding_deployment , input = query_text )
107+ query_vector = embedding ["data" ][0 ]["embedding" ]
96108 else :
97109 query_vector = None
98110
99- # Only keep the text query if the retrieval mode uses text, otherwise drop it
111+ # Only keep the text query if the retrieval mode uses text, otherwise drop it
100112 if not has_text :
101113 query_text = None
102114
103115 # Use semantic L2 reranker if requested and if retrieval mode is text or hybrid (vectors + text)
104116 if overrides .get ("semantic_ranker" ) and has_text :
105- r = await self .search_client .search (query_text ,
106- filter = filter ,
107- query_type = QueryType .SEMANTIC ,
108- query_language = "en-us" ,
109- query_speller = "lexicon" ,
110- semantic_configuration_name = "default" ,
111- top = top ,
112- query_caption = "extractive|highlight-false" if use_semantic_captions else None ,
113- vector = query_vector ,
114- top_k = 50 if query_vector else None ,
115- vector_fields = "embedding" if query_vector else None )
117+ r = await self .search_client .search (
118+ query_text ,
119+ filter = filter ,
120+ query_type = QueryType .SEMANTIC ,
121+ query_language = "en-us" ,
122+ query_speller = "lexicon" ,
123+ semantic_configuration_name = "default" ,
124+ top = top ,
125+ query_caption = "extractive|highlight-false" if use_semantic_captions else None ,
126+ vector = query_vector ,
127+ top_k = 50 if query_vector else None ,
128+ vector_fields = "embedding" if query_vector else None ,
129+ )
116130 else :
117- r = await self .search_client .search (query_text ,
118- filter = filter ,
119- top = top ,
120- vector = query_vector ,
121- top_k = 50 if query_vector else None ,
122- vector_fields = "embedding" if query_vector else None )
131+ r = await self .search_client .search (
132+ query_text ,
133+ filter = filter ,
134+ top = top ,
135+ vector = query_vector ,
136+ top_k = 50 if query_vector else None ,
137+ vector_fields = "embedding" if query_vector else None ,
138+ )
123139 if use_semantic_captions :
124- results = [doc [self .sourcepage_field ] + ": " + nonewlines (" . " .join ([c .text for c in doc ['@search.captions' ]])) async for doc in r ]
140+ results = [
141+ doc [self .sourcepage_field ] + ": " + nonewlines (" . " .join ([c .text for c in doc ["@search.captions" ]]))
142+ async for doc in r
143+ ]
125144 else :
126145 results = [doc [self .sourcepage_field ] + ": " + nonewlines (doc [self .content_field ]) async for doc in r ]
127146 content = "\n " .join (results )
128147
129- follow_up_questions_prompt = self .follow_up_questions_prompt_content if overrides .get ("suggest_followup_questions" ) else ""
148+ follow_up_questions_prompt = (
149+ self .follow_up_questions_prompt_content if overrides .get ("suggest_followup_questions" ) else ""
150+ )
130151
131152 # STEP 3: Generate a contextual and content specific answer using the search results and chat history
132153
133154 # Allow client to replace the entire prompt, or to inject into the exiting prompt using >>>
134155 prompt_override = overrides .get ("prompt_template" )
135156 if prompt_override is None :
136- system_message = self .system_message_chat_conversation .format (injected_prompt = "" , follow_up_questions_prompt = follow_up_questions_prompt )
157+ system_message = self .system_message_chat_conversation .format (
158+ injected_prompt = "" , follow_up_questions_prompt = follow_up_questions_prompt
159+ )
137160 elif prompt_override .startswith (">>>" ):
138- system_message = self .system_message_chat_conversation .format (injected_prompt = prompt_override [3 :] + "\n " , follow_up_questions_prompt = follow_up_questions_prompt )
161+ system_message = self .system_message_chat_conversation .format (
162+ injected_prompt = prompt_override [3 :] + "\n " , follow_up_questions_prompt = follow_up_questions_prompt
163+ )
139164 else :
140165 system_message = prompt_override .format (follow_up_questions_prompt = follow_up_questions_prompt )
141166
142167 messages = self .get_messages_from_history (
143168 system_message ,
144169 self .chatgpt_model ,
145170 history ,
146- history [- 1 ]["user" ]+ "\n \n Sources:\n " + content , # Model does not handle lengthy system messages well. Moving sources to latest user conversation to solve follow up questions prompt.
147- max_tokens = self .chatgpt_token_limit )
148- msg_to_display = '\n \n ' .join ([str (message ) for message in messages ])
149-
150- extra_info = {"data_points" : results , "thoughts" : f"Searched for:<br>{ query_text } + msg_to_display .replace ('\n ' , '<br>' )}
171+ # Model does not handle lengthy system messages well.
172+ # Moved sources to latest user conversation to solve follow up questions prompt.
173+ history [- 1 ]["user" ] + "\n \n Sources:\n " + content ,
174+ max_tokens = self .chatgpt_token_limit ,
175+ )
176+ msg_to_display = "\n \n " .join ([str (message ) for message in messages ])
177+
178+ extra_info = {
179+ "data_points" : results ,
180+ "thoughts" : f"Searched for:<br>{ query_text }
181+ + msg_to_display .replace ("\n " , "<br>" ),
182+ }
151183 chat_coroutine = openai .ChatCompletion .acreate (
152- deployment_id = self .chatgpt_deployment ,
153- model = self .chatgpt_model ,
154- messages = messages ,
155- temperature = overrides .get ("temperature" ) or 0.7 ,
156- max_tokens = 1024 ,
157- n = 1 ,
158- stream = should_stream )
184+ deployment_id = self .chatgpt_deployment ,
185+ model = self .chatgpt_model ,
186+ messages = messages ,
187+ temperature = overrides .get ("temperature" ) or 0.7 ,
188+ max_tokens = 1024 ,
189+ n = 1 ,
190+ stream = should_stream ,
191+ )
159192 return (extra_info , chat_coroutine )
160193
161194 async def run_without_streaming (self , history : list [dict [str , str ]], overrides : dict [str , Any ]) -> dict [str , Any ]:
@@ -164,19 +197,29 @@ async def run_without_streaming(self, history: list[dict[str, str]], overrides:
164197 extra_info ["answer" ] = chat_content
165198 return extra_info
166199
167- async def run_with_streaming (self , history : list [dict [str , str ]], overrides : dict [str , Any ]) -> AsyncGenerator [dict , None ]:
200+ async def run_with_streaming (
201+ self , history : list [dict [str , str ]], overrides : dict [str , Any ]
202+ ) -> AsyncGenerator [dict , None ]:
168203 extra_info , chat_coroutine = await self .run_until_final_call (history , overrides , should_stream = True )
169204 yield extra_info
170205 async for event in await chat_coroutine :
171206 yield event
172207
173-
174- def get_messages_from_history (self , system_prompt : str , model_id : str , history : list [dict [str , str ]], user_conv : str , few_shots = [], max_tokens : int = 4096 ) -> list :
208+ def get_messages_from_history (
209+ self ,
210+ system_prompt : str ,
211+ model_id : str ,
212+ history : list [dict [str , str ]],
213+ user_conv : str ,
214+ few_shots = [],
215+ max_tokens : int = 4096 ,
216+ ) -> list :
175217 message_builder = MessageBuilder (system_prompt , model_id )
176218
177- # Add examples to show the chat what responses we want. It will try to mimic any responses and make sure they match the rules laid out in the system message.
219+ # Add examples to show the chat what responses we want.
220+ # It will try to mimic any responses and make sure they match the rules laid out in the system message.
178221 for shot in few_shots :
179- message_builder .append_message (shot .get (' role' shot .get (' content'
222+ message_builder .append_message (shot .get (" role" shot .get (" content"
180223
181224 user_content = user_conv
182225 append_index = len (few_shots ) + 1
0 commit comments