long-conversation

Long Conversation Context Management

Overview

As conversations grow, they create two problems:

1. Context overflow — the message history exceeds the LLM's context window
2. I/O cost — loading and saving the full history on every turn is expensive

The PromptKit SDK solves both with a three-tier context system:

┌─────────────────────────────────────────────┐
│              Context Assembly                │
│                                              │
│  ┌──────────┐  ┌──────────┐  ┌───────────┐  │
│  │ Summaries│→ │ Retrieved│→ │Hot Window  │  │
│  │ (oldest) │  │ (relevant│  │ (recent N) │  │
│  │          │  │  older)  │  │            │  │
│  └──────────┘  └──────────┘  └───────────┘  │
│                                              │
│  Sent to LLM: summaries + retrieved + recent │
└─────────────────────────────────────────────┘

• Hot window: Only the last N messages are loaded from the store each turn
• Semantic retrieval: Older messages relevant to the current query are found via embeddings
• Auto-summarization: When message count exceeds a threshold, old turns are compressed into summaries

Features Demonstrated

• WithContextWindow(n) — load only the N most recent messages
• WithContextRetrieval(embProvider, topK) — semantic search for relevant older messages
• WithAutoSummarize(provider, threshold, batchSize) — compress old turns into summaries
• WithStateStore(store) — persistent conversation storage
• WithConversationID(id) — explicit session identity for state persistence

Prerequisites

• Go 1.26+
• OpenAI API key (for the LLM, embeddings, and summarization)

Running the Example

export OPENAI_API_KEY=sk-...
go run .

How It Works

Context Window

sdk.WithContextWindow(4)

Instead of loading the entire conversation from the state store, only the last 4 messages are loaded. This uses the ContextAssemblyStage pipeline stage internally. When the store implements MessageReader (as both MemoryStore and RedisStore do), only the tail is deserialized — no full-state load required.

New messages are saved incrementally via IncrementalSaveStage, which uses MessageAppender when available.

Semantic Retrieval

embProvider, _ := openai.NewEmbeddingProvider()
sdk.WithContextRetrieval(embProvider, 3)

On each turn, the user's message is embedded and compared against older messages (those outside the hot window). The top 3 most semantically similar messages are retrieved and inserted chronologically between summaries and the hot window.

This ensures that when a user says "going back to the billing issue," the relevant earlier billing messages are included even though they've scrolled out of the hot window.

Auto-Summarization

summaryProvider := openai.NewProvider(
    "summarizer", "gpt-4o-mini", "https://api.openai.com/v1",
    providers.ProviderDefaults{MaxTokens: 1024}, false,
)
sdk.WithAutoSummarize(summaryProvider, 6, 4)

When the total message count exceeds 6, the 4 oldest unsummarized messages are compressed into a summary using the specified LLM provider. Summaries are prepended to the context as system messages, preserving key information while reducing token count.

Using a smaller model like gpt-4o-mini for summarization keeps costs low while the primary conversation can use a more capable model.

Configuration Reference

Internal Pipeline

When context management options are set, the SDK inserts two pipeline stages:

1. ContextAssemblyStage — runs before the LLM call. Loads summaries, retrieves relevant messages, and loads the hot window. Assembles them into the context sent to the provider.
2. IncrementalSaveStage — runs after the LLM call. Appends only the new messages (user + assistant) to the store, and triggers summarization when the threshold is exceeded.

Production Tips

• Use Redis for distributed, persistent state: statestore.NewRedisStore(redisClient). Redis Lists give O(1) append and tail reads.
• Choose the right embedding model: text-embedding-3-small is fast and cheap; text-embedding-3-large is more accurate. Voyage AI (voyage-3.5) is recommended for Claude-based systems.
• Tune the context window to your use case: 10-20 messages is typical for support; 5-10 for focused Q&A.
• Set summarization threshold based on your model's context size and average message length.
• Use a cheap summarizer: gpt-4o-mini or a similar small model works well for generating summaries.

See Also

• Manage Context How-To — SDK context management options
• State Store Reference — Store interfaces and implementations