smart_messages

Smart Message Merging Example

Background

In modern LLM applications, especially those involving tool use (Function Calling), managing the message history is critical. A common pattern involves an LLM generating a "tool call" message, followed by the tool execution result. Sometimes, we want to update a previous message (e.g., replacing a "Thinking..." placeholder with a final answer) rather than appending a new one. Standard list appending is insufficient for this. The AddMessages Reducer provides intelligent merging logic.

Features

• ID-Based Deduplication: If a new message has the same ID as an existing one, it updates (overwrites) the existing message instead of appending.
• Automatic Appending: Messages without IDs (or with new IDs) are appended to the end of the list.
• Flexibility: Works with both llms.MessageContent (standard) and custom map-based message structures.

Implementation Principle

The AddMessages function in graph/add_messages.go implements the reducer logic:

1. It iterates through the new messages.
2. For each message, it attempts to extract an ID (via MessageWithID interface, map key "id", or struct field "ID").
3. If an ID is found and matches an existing message in the current state list, the existing message is replaced at its original index.
4. Otherwise, the new message is appended.

Code Walkthrough

In main.go:

1. Graph Setup:

   g := graph.NewMessagesStateGraph()
   

   This helper creates a graph where the "messages" key is pre-configured to use AddMessages.

2. AI Response Node: Returns a message with id: "msg_123" and content "Thinking...".

3. AI Update Node: Returns a message with the same id: "msg_123" but content "Hello...".

4. Result: Because the IDs match, the final history contains only one AI message (the updated one), not two.

How to Run

go run main.go