make sys prompt configurable

app.py

@@ -22,9 +22,9 @@ rag = RAG(
 )
 
 
-def generate(prompt, model_parameters):
+def generate(prompt, sys_prompt, model_parameters):
     try:
-        output, context, source = rag.get_response(prompt, model_parameters)
+        output, context, source = rag.get_response(prompt, sys_prompt, model_parameters)
         return output, context, source
     except HTTPError as err:
         if err.code == 400:
@@ -37,7 +37,7 @@ def generate(prompt, model_parameters):
         )
 
 
-def submit_input(input_, num_chunks, max_new_tokens, repetition_penalty, top_k, top_p, do_sample, temperature):
+def submit_input(input_, sysprompt_, num_chunks, max_new_tokens, repetition_penalty, top_k, top_p, do_sample, temperature):
     if input_.strip() == "":
         gr.Warning("Not possible to inference an empty input")
         return None
@@ -53,7 +53,7 @@ def submit_input(input_, num_chunks, max_new_tokens, repetition_penalty, top_k,
         "temperature": temperature
     }
 
-    output, context, source = generate(input_, model_parameters)
+    output, context, source = generate(input_, sysprompt_, model_parameters)
     sources_markup = ""
 
     for url in source:
@@ -112,6 +112,12 @@ def gradio_app():
                     placeholder="Quina és la finalitat del Servei Meteorològic de Catalunya?",
                     # value = "Quina és la finalitat del Servei Meteorològic de Catalunya?"
                 )
+                sysprompt_ = Textbox(
+                    lines=2,
+                    label="System",
+                    placeholder="Below is a question that you should answer based on the given context. Write a response that answers the question using only information provided in the context.",
+                    value = "Below is a question that you should answer based on the given context. Write a response that answers the question using only information provided in the context."
+                )
                 with gr.Row(variant="panel"):
                     clear_btn = Button(
                         "Clear",
@@ -201,8 +207,8 @@ def gradio_app():
             inputs=[input_],
             api_name=False,
             js="""(i, m) => {
-            document.getElementById('inputlenght').textContent = i.length + '  '
-            document.getElementById('inputlenght').style.color =  (i.length > m) ? "#ef4444" : "";
+            document.getElementById('inputlength').textContent = i.length + '  '
+            document.getElementById('inputlength').style.color =  (i.length > m) ? "#ef4444" : "";
         }""",
         )
 
@@ -216,7 +222,7 @@ def gradio_app():
         
         submit_btn.click(
             fn=submit_input, 
-            inputs=[input_]+ parameters_compontents, 
+            inputs=[input_, sysprompt_]+ parameters_compontents, 
             outputs=[output, source_context, context_evaluation],
             api_name="get-results"
         )

rag.py

@@ -32,7 +32,7 @@ class RAG:
 
         return documentos
         
-    def predict(self, instruction, context, model_parameters):
+    def predict(self, instruction, sys_prompt, context, model_parameters):
 
         from openai import OpenAI
         
@@ -42,9 +42,10 @@ class RAG:
             api_key=os.getenv("HF_TOKEN")
         )
 
-        sys_prompt = "You are a helpful assistant. Answer the question using only the context you are provided with. If it is not possible to do it with the context, just say 'I can't answer'. <|endoftext|>"
+        #sys_prompt = "You are a helpful assistant. Answer the question using only the context you are provided with. If it is not possible to do it with the context, just say 'I can't answer'. <|endoftext|>"
         #query = f"Context:\n{context}\n\nQuestion:\n{instruction}"
         query = f"{sys_prompt}\n\nContext:\n{context}\n\nQuestion:\n{instruction}"
+        print(query)
         #query = f"{sys_prompt}\n\nQuestion:\n{instruction}\n\nContext:\n{context}"
         chat_completion = client.chat.completions.create(
             model="tgi",
@@ -77,14 +78,14 @@ class RAG:
 
         return text_context, full_context, source_context
 
-    def get_response(self, prompt: str, model_parameters: dict) -> str:
+    def get_response(self, prompt: str, sys_prompt: str, model_parameters: dict) -> str:
         try:
             docs = self.get_context(prompt, model_parameters["NUM_CHUNKS"])
             text_context, full_context, source = self.beautiful_context(docs)
 
             del model_parameters["NUM_CHUNKS"]
 
-            response = self.predict(prompt, text_context, model_parameters)
+            response = self.predict(prompt, sys_prompt, text_context, model_parameters)
 
             if not response:
                 return self.NO_ANSWER_MESSAGE