minor improvements in llm prompts

app.py

@@ -23,7 +23,7 @@ doc_content_fr = Doc(content_fr_path)
 client_db = chromadb.Client()
 retriever = Retriever(client_db, doc_plan, doc_content, doc_content_fr, collection_name)
 
-llm_model = OpenAI(temperature=0)
+llm_model = OpenAI(temperature=0, model_name="gpt-3.5-turbo")
 llm = LlmAgent(llm_model)
 
 specials['remote_rate_df'] = pd.read_csv(specials['remote_rate_path'])

config.py

@@ -25,19 +25,19 @@ view_config = {
 }
 
 countries_extensions = {
-    'Royaume-Uni': ['UK', 'U.K.','RU', 'R.U.', 'Angleterre'],
-    'Etats-Unis': ['Etats-unis', 'Etats Unis', 'Etats unis', 'ETATS-UNIS', 'USA'],
+    'Grande Bretagne': ['UK', 'U.K.', 'RU', 'R.U.', 'Angleterre', 'Grande-Bretagne', 'Royaume-Uni', 'Royaume Uni',
+                        "l'Angleterre", 'Grande Bretagne'],
+    'Etats-Unis': ['Etats-unis', 'Etats Unis', 'Etats unis', 'ETATS-UNIS', 'USA', 'U.S.A', 'U.S.A.'],
     'E.A.U': ["EAU", "Emirats", "Emirats Arabes Unis", "Emirates", "UAE", "United Arab Emirates"],
     'Pays-Bas': ['Les Pays-Bas', 'Hollande', 'Holland']
 }
 specials = {'remote_rate_path': 'data/remote_rates.csv',
             'remote_rate_known': "the scale rate of remoteness for the ",
-
-            'remote_rate_unknown': "the scale rate of remoteness for the country  mentionned is unknown. Allowances "
+            'remote_rate_unknown': "the scale rate of remoteness for the mentioned country is unknown. Allowances "
             "apply though",
             'accommodation_meal_path': 'data/accommodation_meal_rates.csv',
             'accommodation_meal_known': 'the rates for accommodation and meals are the following: ',
-            'accommodation_meal_unknown': 'the rates for accommodation and meals are not defined for the country '
-                                              'mentionned ',
+            'accommodation_meal_unknown':
+                'the rates for accommodation and meals are not defined for the mentioned country ',
             'countries_extensions': countries_extensions,
             }

src/control/control.py

@@ -28,12 +28,13 @@ class Controller:
         sources_contents_fr = [s.content_fr for s in block_sources[:2]]
         context_fr = '\n'.join(sources_contents_fr)
         if self.content_language == 'en':
-            answer = self.llm.generate_answer(answer_en=answer, query=query_fr, context_fr=context_fr)
-        answer = answer.strip().strip("'''").strip("```")
+            answer = self.llm.generate_answer(answer_en=answer, query=query_fr,
+                                              histo_fr=histo_conversation, context_fr=context_fr)
+        answer = self._clean_answer(answer)
         return answer, block_sources
 
     @staticmethod
-    def _get_histo(histo: [(str, str)]) -> str:
+    def _get_histo(histo: [(str, str)]) -> (str, str):
         histo_conversation = ""
         histo_queries = ""
 
@@ -42,6 +43,17 @@ class Controller:
             histo_queries += query + '\n'
         return histo_conversation[:-1], histo_queries
 
+    @staticmethod
+    def _clean_answer(answer: str) -> str:
+        answer = answer.strip('bot:')
+        while answer and answer[-1] in {"'", '"', " ", "`"}:
+            answer = answer[:-1]
+        while answer and answer[0] in {"'", '"', " ", "`"}:
+            answer = answer[1:]
+        answer = answer.strip('bot:')
+        if answer and answer[-1] != ".":
+            answer += "."
+        return answer
 
     @staticmethod
     def _select_best_sources(sources: [Block], delta_1_2=0.15, delta_1_n=0.3, absolute=1.2, alpha=0.9) -> [Block]:

src/tools/llm.py

@@ -5,14 +5,15 @@ class LlmAgent:
 
     def generate_paragraph(self, query: str, context: {}, histo: [(str, str)], language='fr') -> str:
         """generates the  answer"""
-        template = (f"You are a bot designed to answer to the query from users delimited by triple backticks: \\n"
-                    f"``` {query} ```\\n"
-                    f"Your answer is based on the context delimited by triple backticks: \\n"
-                    f"``` {context} ```\\n"
-                    f"Your answer takes into account previous part of the conversation in french delimited by triple "
-                    f"backticks and shall avoid redundancies in the answers\\n "
-                    f"``` {histo} ```\\n"
-                    f"Your response shall be in {language} and shall be concise/"
+        template = (f"You are a conversation bot designed to answer to the query from users delimited by "
+                    f"triple backticks: "
+                    f"\\n ``` {query} ```\\n"
+                    f"Your answer is based on the context delimited by triple backticks: "
+                    f"\\n ``` {context} ```\\n"
+                    f"You are consistent and avoid redundancies with the rest of the initial conversation in french"
+                    f"delimited by triple backticks: "
+                    f"\\n ``` {histo} ```\\n"
+                    f"Your response shall be in {language} and shall be concise"
                     f"In case the provided context is not relevant to answer to the question, just return that you "
                     f"don't know the answer ")
 
@@ -36,16 +37,20 @@ class LlmAgent:
         p = self.llm(template)
         return p
 
-    def generate_answer(self, query: str, answer_en: str, context_fr: str) -> str:
+    def generate_answer(self, query: str, answer_en: str, histo_fr: str, context_fr: str) -> str:
         """provides the final answer in french based on the initial query and the answer in english"""
 
         def _cut_unfinished_sentence(s: str):
             return '.'.join(p.split('.')[:-1])
 
-        template = (f"Your task consists in providing the answer in french to the query "
-                    f"delimited by triple backticks: ```{query}``` given the informations here delimited "
-                    f"by triple backticks: ```{context_fr}``` and the answer in english delimited by triple "
-                    f"backticks: ```{answer_en}```"
+        template = (f"Your task consists in translating the answer in french to the query "
+                    f"delimited by triple backticks: ```{query}``` \\n"
+                    f"You are given the answer in english delimited by triple backticks: ```{answer_en}```"
+                    f"\\n You don't add new content to the answer in english but: "
+                    f"\\n 1 You can use some vocabulary from the context in french delimited by triple backticks: "
+                    f"```{context_fr}```"
+                    f"\\n 2 You are consistent and avoid redundancies with the rest of the initial"
+                    f" conversation in french delimited by triple backticks: ```{histo_fr}```"
                     )
 
         p = self.llm(template)

src/view/view.py

@@ -45,7 +45,7 @@ def run(ctrl: Controller, config: {}):
             histo_text_.append((input_text_, None))
             update_ = {
                 histo_text_comp: gr.update(visible=True, value=histo_text_),
-                input_example_comp: gr.update(visible=False, interactive=False),
+                input_example_comp: gr.update(visible=False,),
             }
             for i in range(4):
                 update_[source_text_comp[i]] = gr.update(visible=False)
@@ -70,7 +70,7 @@ def run(ctrl: Controller, config: {}):
             update_ = {
                 input_text_comp: gr.update(value=input_example_),
                 histo_text_comp: gr.update(visible=True, value=histo_text_),
-                input_example_comp: gr.update(visible=False, interactive=False, value=''),
+                input_example_comp: gr.update(visible=False, value=''),
             }
             for i in range(4):
                 update_[source_text_comp[i]] = gr.update(visible=False)
@@ -80,7 +80,7 @@ def run(ctrl: Controller, config: {}):
             update_ = {
                 input_text_comp: gr.update(value=''),
                 histo_text_comp: gr.update(value='', visible=False),
-                input_example_comp: gr.update(value='', visible=True, interactive=True),
+                input_example_comp: gr.update(value='', visible=True),
             }
             for i in range(4):
                 update_[source_text_comp[i]] = gr.update(visible=False, value='hello')