feat: separando as etapas de gerar relatório e gerar resumo

_utils/resumo_completo_cursor.py

@@ -317,15 +317,16 @@ class ContextualRetriever:
         return contextualized_chunks
 
 class EnhancedDocumentSummarizer(DocumentSummarizer):
-    def __init__(self, openai_api_key: str, claude_api_key: str, config: RetrievalConfig, embedding_model, chunk_size, chunk_overlap, num_k_rerank, model_cohere_rerank, claude_context_model, system_prompt, gpt_model, gpt_temperature, id_modelo_do_usuario):
+    def __init__(self, openai_api_key: str, claude_api_key: str, config: RetrievalConfig, embedding_model, chunk_size, chunk_overlap, num_k_rerank, model_cohere_rerank, claude_context_model, prompt_relatorio, gpt_model, gpt_temperature, id_modelo_do_usuario, prompt_modelo):
         super().__init__(openai_api_key, os.environ.get("COHERE_API_KEY"), embedding_model, chunk_size, chunk_overlap, num_k_rerank, model_cohere_rerank)
         self.config = config
         self.contextual_retriever = ContextualRetriever(config, claude_api_key, claude_context_model)
         self.logger = logging.getLogger(__name__)
-        self.system_prompt = system_prompt
+        self.prompt_relatorio = prompt_relatorio
         self.gpt_model = gpt_model
         self.gpt_temperature = gpt_temperature
         self.id_modelo_do_usuario = id_modelo_do_usuario
+        self.prompt_modelo = prompt_modelo
 
     def create_enhanced_vector_store(self, chunks: List[ContextualizedChunk]) -> Tuple[Chroma, BM25Okapi, List[str]]:
         """Create vector store and BM25 index with contextualized chunks"""
@@ -453,8 +454,6 @@ class EnhancedDocumentSummarizer(DocumentSummarizer):
                         'relevance_score': score,
                         'context': metadata.get('context', '')
                     })
-
-            prompt_template = self.system_prompt
             
             url_request = f"{api_url}/modelo/{self.id_modelo_do_usuario}"
             resposta = requests.get(url_request)
@@ -464,11 +463,6 @@ class EnhancedDocumentSummarizer(DocumentSummarizer):
             
             modelo_buscado = resposta.json()["modelo"]
             
-            prompt = PromptTemplate(
-                template=prompt_template,
-                input_variables=["context", "modelo_usuario"]
-            )
-            
             llm = ChatOpenAI(
                 temperature=self.gpt_temperature,
                 model_name=self.gpt_model, 
@@ -476,10 +470,22 @@ class EnhancedDocumentSummarizer(DocumentSummarizer):
                 
             )
             
-            response = llm.predict(prompt.format(context="\n\n".join(contexts), modelo_usuario=modelo_buscado))
+            prompt_gerar_relatorio = PromptTemplate(
+                template=self.prompt_relatorio,
+                input_variables=["context"]
+            )
+            
+            relatorio_gerado = llm.predict(prompt_gerar_relatorio.format(context="\n\n".join(contexts)))
+
+            prompt_gerar_modelo = PromptTemplate(
+                template=self.prompt_modelo,
+                input_variables=["context", "modelo_usuario"]
+            )
+            
+            modelo_gerado = llm.predict(prompt_gerar_modelo.format(context=relatorio_gerado, modelo_usuario=modelo_buscado))
             
             # Split the response into paragraphs
-            summaries = [p.strip() for p in response.split('\n\n') if p.strip()]
+            summaries = [p.strip() for p in modelo_gerado.split('\n\n') if p.strip()]
             
             # Create structured output
             structured_output = []
@@ -525,10 +531,11 @@ def get_llm_summary_answer_by_cursor_complete(serializer, listaPDFs):
         num_k_rerank=serializer["num_k_rerank"],
         model_cohere_rerank=serializer["model_cohere_rerank"],
         claude_context_model=serializer["claude_context_model"],
-        system_prompt=serializer["system_prompt"],
+        prompt_relatorio=serializer["prompt_relatorio"],
         gpt_model=serializer["model"],
         gpt_temperature=serializer["gpt_temperature"],
-        id_modelo_do_usuario=serializer["id_modelo_do_usuario"]
+        id_modelo_do_usuario=serializer["id_modelo_do_usuario"],
+        prompt_modelo=serializer["prompt_modelo"]
     )
 
     # # Load and process document
@@ -582,5 +589,6 @@ def get_llm_summary_answer_by_cursor_complete(serializer, listaPDFs):
         "hf_embedding": serializer["hf_embedding"],
         "chunk_size": serializer["chunk_size"],
         "chunk_overlap": serializer["chunk_overlap"],
-        "system_prompt": serializer["system_prompt"],
+        "prompt_relatorio": serializer["prompt_relatorio"],
+        "prompt_modelo": serializer["prompt_modelo"]
     }}

resumos/serializer.py

@@ -29,7 +29,20 @@ class ResumoCursorSerializer(serializers.Serializer):
     chunk_overlap = serializers.IntegerField(required=False, default=200)
 
 
-system_prompt = """
+system_prompt_relatorio = """
+            Based on the following context, provide multiple key points from the document.
+            For each point, create a new paragraph.
+            Each paragraph should be a complete, self-contained insight.
+            Include any relevant context provided.
+            
+            Context: {context}
+            
+            Key points:
+            """
+
+user_message = "What are the main points of this document?"
+
+system_prompt_modelo = """
             Based on the following context, provide multiple key points from the document.
             For each point, create a new paragraph.
             Each paragraph should be a complete, self-contained insight.
@@ -41,9 +54,10 @@ system_prompt = """
             
             Key points:
             """
-user_message = "What are the main points of this document?"
 class ResumoCursorCompeltoSerializer(ResumoCursorSerializer):
-    system_prompt = serializers.CharField(required=False, default=system_prompt)
+    system_prompt = None
+    prompt_relatorio = serializers.CharField(required=False, default=system_prompt_relatorio)
+    prompt_modelo = serializers.CharField(required=False, default=system_prompt_modelo)
     user_message = serializers.CharField(required=False, default=user_message)
     num_chunks_retrieval = serializers.IntegerField(default=5)
     embedding_weight = serializers.FloatField(default=0.5)