maybe

.gitignore

@@ -1,8 +1,9 @@
 config_key.py
 
 
-#Test folder
+#Test folder + files
 data/Test/
+test.py
 
 #database folder
 database/

src/tools/llm.py

@@ -16,15 +16,16 @@ class LlmAgent:
     def generate_paragraph(self, query: str, context: {}, histo: [(str, str)], language='fr') -> str:
         locallm = HuggingFacePipeline(pipeline=self.pipe)
         """generates the  answer"""
-        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 "
-                    f"delimited by triple backticks: "
-                    f"\\n ``` {histo} ```\\n"
-                    f"Your response shall be in {language} and shall be concise")
+        template = (f'''[INST] <<SYS>>"
+                    "You are a conversation bot designed to answer to the query from users"
+                    "Your answer is based on the context delimited by triple backticks: "
+                    "\\n ``` {context} ```\\n"
+                    " You are consistent and avoid redundancies with the rest of the initial conversation "
+                    "delimited by triple backticks: "
+                    "\\n ``` {histo} ```\\n"
+                    "Your response shall be in {language} and shall be concise"
+                    "\\n <</SYS>>"
+                    "\\n {query}[/INST]''')
         prompt = PromptTemplate(input_variables=[], template=template)
         llm_chain = LLMChain(prompt=prompt,llm=locallm)
         p = llm_chain.predict()
@@ -40,8 +41,11 @@ class LlmAgent:
 
         # languages = "`French to English" if language == "en" else "English to French"
 
-        tempate = (f"    Your task consists in translating in English\\n"
-                    f"    the following text delimited by by triple backticks: ```{text}```\n"
+        tempate = (f'''[INST] <<SYS>>
+                   Your task consists in translating in English\\n"
+                    the following text:
+                    <</SYS>>
+                    {text}[/INST]'''
                     )
 
         prompt = PromptTemplate(input_variables=[], template=tempate)
@@ -54,14 +58,17 @@ class LlmAgent:
         def _cut_unfinished_sentence(s: str):
             return '.'.join(s.split('.')[:-1])
         locallm = HuggingFacePipeline(pipeline=self.pipe)
-        template = (f"Your task consists in translating the answer in {language}, if its not already the case, to the query "
-                    f"delimited by triple backticks: ```{query}``` \\n"
-                    f"You are given the answer in {language} delimited by triple backticks: ```{answer}```"
-                    f"\\n You don't add new content to the answer but: "
-                    f"\\n 1 You can use some vocabulary from the context delimited by triple backticks: "
-                    f"```{context}```"
-                    f"\\n 2 You are consistent and avoid redundancies with the rest of the initial"
-                    f" conversation delimited by triple backticks: ```{histo}```"
+        template = (f'''[INST] <<SYS>>
+                    Your task consists in translating the answer in {language}, if its not already the case, to the query "
+                    delimited by triple backticks: ```{query}``` \\n"
+                    \\n You don't add new content to the answer but: "
+                    \\n 1 You can use some vocabulary from the context delimited by triple backticks: "
+                    ```{context}```"
+                    \\n 2 You are consistent and avoid redundancies with the rest of the initial"
+                    conversation delimited by triple backticks: ```{histo}```"
+                    You are given the answer in {language}:
+                    <</SYS>>
+                    {answer}[/INST]'''
                     )
         prompt = PromptTemplate(input_variables=[], template=template)
         llm_chain = LLMChain(prompt=prompt,llm=locallm,verbose=True)
@@ -101,9 +108,11 @@ class LlmAgent:
     def detect_language(self, text: str) -> str:
         """detects the language"""
         locallm = HuggingFacePipeline(pipeline=self.pipe)
-        template = (f"Your task consists in detecting the language of the following text delimited by triple backticks: "
-                    f"```{text}```"
-                    f" Your answer shall be the two letters code of the language"
+        template = (f'''[INST] <<SYS>>
+                    Your task consists in detecting the language of the user query"
+                    Your answer shall be the two letters code of the language"
+                    \\n <</SYS>>"
+                    \\n {text}[/INST]'''
                     )
         prompt = PromptTemplate(input_variables=[], template=template)
         llm_chain = LLMChain(prompt=prompt,llm=locallm,verbose=True)

test.py

@@ -1,7 +0,0 @@
-from src.model.doc import Doc
-from config import *
-from src.tools.llm import LlmAgent
-
-llmagent = LlmAgent(model="TheBloke/Llama-2-7b-Chat-GPTQ")
-doc = Doc(path=content_en_path_real)
-[llmagent.transform_parahraph_into_question(block.content, title_doc=doc.title,title_para=block.title) for block in doc.blocks]