会话管理与上下文持久化
学习在多次交互之间维护对话状态和用户上下文，打造流畅的 LLM 体验。
会话管理与上下文持久化 是 CoddyKit 上的免费 LLM Apps in Production (RAG + Vector DB + Caching) 课时。 这是第 2 节课，共 4 节。 你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。 这是 LLM Apps in Production (RAG + Vector DB + Caching) 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 LLM Apps in Production (RAG + Vector DB + Caching) 课程共包含 4 节课。
本课时的部分内容尚未翻译，以英文显示。
Why LLMs Need Memory
Imagine talking to someone who forgets everything you said a moment ago. That's often how Large Language Models (LLMs) work by default!
For a truly natural and helpful experience, LLM applications need to remember past interactions. This is where session management and context persistence come in.
LLMs: Stateless by Design
When you send a prompt to an LLM API, it processes that single request independently. It doesn't inherently 'remember' any previous prompts or responses.
- Each API call is a fresh start.
- This stateless nature is efficient for simple, one-off questions.
- But it breaks down for conversations or personalized tasks.
Keeping the Conversation Flow
Context persistence is the technique of storing and retrieving relevant past information to include with new LLM requests.
This allows the LLM to understand the ongoing conversation, user preferences, or specific details provided earlier, making its responses much more coherent and useful.
Basic Strategy: Conversation History
The most common way to persist context for chat-based LLM applications is to maintain a conversation history.
- Each user query and LLM response is added to a list.
- Before sending a new user query, this entire history is included in the prompt.
- This gives the LLM the full 'memory' of the interaction.
Simulating Chat History
Let's see a simple Python example where we build up a conversation history in a list. Notice how new messages are appended.
def simulate_chat():
chat_history = []
chat_history.append({"role": "user", "content": "Hi there!"})
chat_history.append({"role": "assistant", "content": "Hello! How can I help?"})
chat_history.append({"role": "user", "content": "What's the weather?"})
print("--- Current Chat History ---")
for msg in chat_history:
print(f"{msg['role']}: {msg['content']}")
if __name__ == "__main__":
simulate_chat()Limitations of In-Memory History
While simple Python lists are great for demonstration, they have big limitations for real-world apps:
- Ephemeral: Data is lost if the application restarts.
- Single Session: Only works for one user's current interaction.
- Scaling Issues: Not suitable for multiple concurrent users.
We need more robust solutions for persistence!
Storing Context Externally
To overcome in-memory limitations, context must be stored in an external, persistent system.
Common choices include:
- Databases: SQL (PostgreSQL, MySQL) or NoSQL (MongoDB, Cassandra) for structured history.
- Key-Value Stores: Redis or Memcached for fast access to session data.
- Cloud Storage: Object storage like S3 for less frequent access.
Context in Action: LLM Call
When using external storage, the process looks like this:
- User sends a new message.
- Application retrieves the user's past conversation context from the external store.
- The full context (history + new message) is sent to the LLM.
- LLM generates a response.
- The new response is added to the context and saved back to the external store.
Conceptual Code: Using Stored Context
This conceptual snippet shows how you'd load history and combine it with a new message before sending to an LLM. Assume load_history() and save_history() interact with an external store.
def send_to_llm_with_context(user_id, new_message):
# Imagine these load/save from Redis/DB
def load_history(uid): return [] # Placeholder
def save_history(uid, hist): pass # Placeholder
history = load_history(user_id)
history.append({"role": "user", "content": new_message})
# Construct the full prompt for the LLM
llm_prompt = "".join([f"{msg['role']}: {msg['content']}\n" for msg in history])
llm_prompt += "Assistant: "
print(f"--- Sending to LLM ---\n{llm_prompt}")
# Simulate LLM response
llm_response = "I understand."
history.append({"role": "assistant", "content": llm_response})
save_history(user_id, history)
if __name__ == "__main__":
send_to_llm_with_context("user_123", "Tell me about context persistence.")More Than Just Chat History
Context persistence isn't limited to just conversation history. It can also include:
- User Profiles: Name, preferences, location.
- Application State: Current task, active selections.
- Document References: Which documents a user has interacted with.
This enriches the LLM's understanding and allows for truly personalized experiences.
Check Your Understanding
Understanding why LLMs need context is crucial for building robust applications.
Recap: Remembering the Past
In this lesson, we explored the critical role of session management and context persistence for LLM applications.
- LLMs are stateless, requiring explicit context.
- Conversation history is a primary form of context.
- External storage (databases, Redis) is vital for robust persistence.
- Context goes beyond chat, including user profiles and app state.
Mastering context persistence is key to creating intuitive and powerful LLM experiences!
常见问题解答
「会话管理与上下文持久化」课时是免费的吗？
是的 — 「会话管理与上下文持久化」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 LLM Apps in Production (RAG + Vector DB + Caching) 课程的其余内容，请升级到 CoddyKit PRO。 LLM Apps in Production (RAG + Vector DB + Caching) 课程共包含 4 节课。
「会话管理与上下文持久化」这节课中我会学到什么？
学习在多次交互之间维护对话状态和用户上下文，打造流畅的 LLM 体验。 你通过在浏览器中直接运行的动手代码来练习 LLM Apps in Production (RAG + Vector DB + Caching)，全天候 AI 导师会在你学习这节课的过程中回答你的问题。
学习 LLM Apps in Production (RAG + Vector DB + Caching) 需要有经验吗？
无需任何先前经验。CoddyKit 上的 LLM Apps in Production (RAG + Vector DB + Caching) 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。 这是第 2 节课，共 4 节。
「会话管理与上下文持久化」课时需要多长时间？
大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。
我能在这节 LLM Apps in Production (RAG + Vector DB + Caching) 课中编写并运行代码吗？
能。每节 LLM Apps in Production (RAG + Vector DB + Caching) 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。