Method insert_pages_into_session

memoryos-pypi/mid_term.py:188–277

(self, summary_for_new_pages, keywords_for_new_pages, pages_to_insert, 
                                  similarity_threshold=0.6, keyword_similarity_alpha=1.0)

        # No save here, it's an internal operation often followed by other ops that save

    def insert_pages_into_session(self, summary_for_new_pages, keywords_for_new_pages, pages_to_insert, 
                                  similarity_threshold=0.6, keyword_similarity_alpha=1.0):
        if not self.sessions: # If no existing sessions, just add as a new one
            print("MidTermMemory: No existing sessions. Adding new session directly.")
            return self.add_session(summary_for_new_pages, pages_to_insert, keywords_for_new_pages)

        new_summary_vec = get_embedding(
            summary_for_new_pages,
            model_name=self.embedding_model_name,
            **self.embedding_model_kwargs
        )
        new_summary_vec = normalize_vector(new_summary_vec)
        
        best_sid = None
        best_overall_score = -1

        for sid, existing_session in self.sessions.items():
            existing_summary_vec = np.array(existing_session["summary_embedding"], dtype=np.float32)
            semantic_sim = float(np.dot(existing_summary_vec, new_summary_vec))
            
            # Keyword similarity (Jaccard index based)
            existing_keywords = set(existing_session.get("summary_keywords", []))
            new_keywords_set = set(keywords_for_new_pages)
            s_topic_keywords = 0
            if existing_keywords and new_keywords_set:
                intersection = len(existing_keywords.intersection(new_keywords_set))
                union = len(existing_keywords.union(new_keywords_set))
                if union > 0:
                    s_topic_keywords = intersection / union 
            
            overall_score = semantic_sim + keyword_similarity_alpha * s_topic_keywords
            
            if overall_score > best_overall_score:
                best_overall_score = overall_score
                best_sid = sid
        
        if best_sid and best_overall_score >= similarity_threshold:
            print(f"MidTermMemory: Merging pages into session {best_sid}. Score: {best_overall_score:.2f} (Threshold: {similarity_threshold})")
            target_session = self.sessions[best_sid]
            
            processed_new_pages = []
            for page_data in pages_to_insert:
                page_id = page_data.get("page_id", generate_id("page")) # Use existing or generate new ID
                
                # 检查是否已有embedding，避免重复计算
                if "page_embedding" in page_data and page_data["page_embedding"]:
                    print(f"MidTermMemory: Reusing existing embedding for page {page_id}")
                    inp_vec = page_data["page_embedding"]
                    # 确保embedding是normalized的
                    if isinstance(inp_vec, list):
                        inp_vec_np = np.array(inp_vec, dtype=np.float32)
                        if np.linalg.norm(inp_vec_np) > 1.1 or np.linalg.norm(inp_vec_np) < 0.9:  # 检查是否需要重新normalize
                            inp_vec = normalize_vector(inp_vec_np).tolist()
                else:
                    print(f"MidTermMemory: Computing new embedding for page {page_id}")
                    full_text = f"User: {page_data.get('user_input','')} Assistant: {page_data.get('agent_response','')}"
                    inp_vec = get_embedding(
                        full_text,