Multi-turn chat

When you need a sustained conversation and want explicit control over it — for example, one session per user, per request, or per workflow stage — start sessions explicitly with StartNewChatAsync and address them by ID.

When to use this pattern

  • Multiple users, each with their own conversation on a single process.
  • Multi-stage workflows where each stage has its own dialogue history.
  • Long conversations where you need to clean up the KV cache manually.
  • Any scenario where the current-session shortcut from Simple chat is not enough.

Prerequisites

Start a session

string sessionId = await api.StartNewChatAsync();

Optional overrides:

string sessionId = await api.StartNewChatAsync(
    preset: myPreset,               // use a different preset for this session
    sessionId: "user-42-conv-1");   // your own identifier; uniqueness is your responsibility

Send messages to the session

Use SendMessageToSessionAsync with the session ID. The model sees the full history of the session and answers in context.

string reply1 = await api.SendMessageToSessionAsync(sessionId, "I want to plan a trip to Lisbon in May.");
string reply2 = await api.SendMessageToSessionAsync(sessionId, "Suggest three neighborhoods to stay in.");
string reply3 = await api.SendMessageToSessionAsync(sessionId, "Which one is closest to the beach?");

Every message adds to the session’s KV cache. The third reply can reference the first two turns because the earlier context is preserved.

Full example — single session

using Aspose.LLM;
using Aspose.LLM.Abstractions.Parameters.Presets;

var license = new Aspose.LLM.License();
license.SetLicense("Aspose.LLM.lic");

var preset = new Qwen25Preset();
preset.ChatParameters.SystemPrompt = "You are a concise travel assistant. Answer in one or two sentences.";

using var api = AsposeLLMApi.Create(preset);

string sessionId = await api.StartNewChatAsync();

string[] userMessages =
{
    "I want to plan a trip to Lisbon in May.",
    "Suggest three neighborhoods to stay in.",
    "Which one is closest to the beach?",
    "How long does it take to walk from there to the historic center?",
    "Any recommendations for lunch on the way?",
};

foreach (string message in userMessages)
{
    Console.WriteLine($"> {message}");
    string reply = await api.SendMessageToSessionAsync(sessionId, message);
    Console.WriteLine(reply);
    Console.WriteLine();
}

Each follow-up sees the previous turns, so the model can track the topic across the whole exchange.

Multiple concurrent sessions

A single AsposeLLMApi instance can host many sessions at once. Each session keeps its own history and KV region.

string sessionA = await api.StartNewChatAsync(sessionId: "conv-A");
string sessionB = await api.StartNewChatAsync(sessionId: "conv-B");

await api.SendMessageToSessionAsync(sessionA, "Let us talk about ancient Rome.");
await api.SendMessageToSessionAsync(sessionB, "Let us talk about modern cooking.");

string replyA = await api.SendMessageToSessionAsync(sessionA, "Name three emperors.");
string replyB = await api.SendMessageToSessionAsync(sessionB, "Name three essential knives.");

Console.WriteLine($"A: {replyA}");
Console.WriteLine($"B: {replyB}");

The topics do not mix — sessionA sees only the Rome thread, sessionB only the cooking thread.

Manage KV cache in long sessions

As a session grows, its KV cache approaches the preset’s ContextParameters.ContextSize. When the next message would overflow, the engine trims automatically according to ChatParameters.CacheCleanupStrategy (default RemoveOldestMessages).

You can also trim explicitly — for example, at a natural topic boundary in your application:

api.ForceCacheCleanup(CacheCleanupStrategy.KeepSystemPromptOnly);

This operates on the current session. To trim a different session, set it as current first (e.g., by calling SendMessageToSessionAsync once on that session, which does not change ChatParameters but keeps the session warm), or use the default strategy and let the engine trim when needed.

See Chat sessions — Manage the KV cache for all five strategies and how to pick one.

Pick a session ID strategy

  • Generated IDs — let the engine choose (StartNewChatAsync() with no sessionId). Safe for single-process scenarios where you store the returned ID in your code or database.
  • Your own IDs — pass a meaningful ID like "user-42-conv-1". Makes logs and persisted session files easier to inspect. You are responsible for uniqueness within the process.

Common errors

  • Not licensed for this method — apply a license before starting sessions.
  • Context overflow before trim — if you see nonsensical output in a long session, the model may have lost track of the system prompt. Force an explicit ForceCacheCleanup(KeepSystemPromptAndFirstUserMessage) at topic boundaries.
  • Mixing sessions — double-check you pass the right sessionId to SendMessageToSessionAsync. Using SendMessageAsync in a multi-session app sends to the current session, which may not be the one you expect.

What’s next

Simple chat Save and restore session