Custom Redis extends BaseCheckpointSaver (Error: Unable to coerce message from array)

[witvoet-hamiltonai]

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

async _loadBlobs(blobValues: Record<string, any>) {
    if (!blobValues || Object.keys(blobValues).length === 0) {
      return {};
    }

    const decoder = new TextDecoder();

    const entries = await Promise.all(
      Object.entries(blobValues)
        .filter(([, value]) => {
          return (
            typeof value === "object" &&
            value !== null &&
            value.type !== "empty" // Exclude empty values
          );
        })
        .map(async ([key, value]) => {
          // Decode the key (if it's not a string, use TextDecoder)
          const decodedKey =
            typeof key === "string" ? key : decoder.decode(key);

          // Check if the value has a 'type' field and process it
          if (typeof value === "object" && value.type) {
            const decodedValue = await this.serde.loadsTyped(value.type, value);
            return [decodedKey, decodedValue];
          } else {
            console.warn("Unexpected format, continuing");
            return [decodedKey, value]; // Fallback to returning the raw value
          }
        }),
    );

    const result = Object.fromEntries(entries);
    // console.log("Loaded blobs result:", result);
    return result;
  }

Description

Hi, we're trying to implement our custom RedisCheckPointSaver which extends BaseCheckPointSaver as we need to persist memory for our custom agent we built.

As there was no documentation for this yet we took the PostgresSaver and worked from there.

Functions as expected, expect for when we initiate a tool call and it throws the following:

`Error: Unable to coerce message from array: only human, AI, or system message coercion is currently supported.

Received: {
"lc": 1,
"type": "constructor",
"id": [
"langchain_core",
"messages",
"ToolMessage"
],
"kwargs": {`

I suspect it is something with how we've implemented serialization, example is from our _loadBlobs function.

I have been over all the API documentation, Python and JS examples to pull something together but hoping someone from the community has a pointer.

I have also tried to filter out the ToolMessages, but this results in OpenAI error tool_msg must be followed by tool call etc.

System Info

"@langchain/core": "^0.3.10",
"@langchain/langgraph": "^0.2.14",
"@langchain/langgraph-checkpoint": "^0.0.10",
"@langchain/langgraph-checkpoint-postgres": "^0.0.1",
"@langchain/openai": "^0.3.7",
"@upstash/redis": "^1.32.0",
"langchain": "^0.3.2",
"redis": "^4.7.0",

[dosubot[bot]]

Hello, @witvoet-hamiltonai! I'm here to help you with your issue. I can assist with bug fixes, answer your questions, and guide you to become a contributor. Let's work together to solve the problem!

The error you're encountering suggests that the coerceMessageLikeToMessage function is not handling ToolMessage types correctly. This function currently supports coercion for "human", "AI", and "system" messages, but not for "tool" messages.

To address this, you might need to modify the coerceMessageLikeToMessage function to include support for ToolMessage. Here's a general approach:

1. Identify the Message Type: Ensure that the ToolMessage type is correctly identified in your serialization/deserialization logic.

2. Extend Coercion Logic: Update the coerceMessageLikeToMessage function to handle ToolMessage types. You can add a case to check for ToolMessage and handle it appropriately.

3. Check Serialization: Ensure that your serialization process correctly marks ToolMessage types so they can be identified during deserialization.

4. Review the _loadBlobs Function: Make sure that the serde.loadsTyped method is correctly processing ToolMessage types. You might need to adjust how ToolMessage is deserialized.

By extending the coercion logic to include ToolMessage, you should be able to resolve the error. If you need further details on how ToolMessage is structured, it includes fields like tool_call_id, status, and artifact ^[1].

To continue talking to Dosu, mention @dosu.

[witvoet-hamiltonai]

Thanks for the quick reply! Two follow up questions:

1. By editing the function, do you mean directly in the Langchain node_modules as customization to the current function? Or is there an easier way to "overwrite" this function with custom logic.

2. It seems that on the message arrays the our deserialization function never gets called because the array that contains the messages is not an "object". Perhaps our structure is off in Redis? Note that we save and update the entire chat history of a thread in 1 redis k/v store.

{"checkpoint":{"v":1,"id":"1ef89ccc-29c4-6690-8001-36507e3f14ce","ts":"2024-10-14T01:36:40.313Z","channel_versions":{"__start__":2,"messages":3,"__start__:agent":3,"agent":3},"versions_seen":{"__input__":{},"__start__":{"__start__":1},"agent":{"__start__:agent":2}},"pending_sends":[]},"channel_values":{"messages":[{"lc":1,"type":"constructor","id":["langchain_core","messages","HumanMessage"],"kwargs":{"content":"test","additional_kwargs":{},"response_metadata":{}}},{"lc":1,"type":"constructor","id":["langchain_core","messages","AIMessageChunk"],"kwargs":{"content":"It looks like you're testing the system. How can I assist you today?","additional_kwargs":{},"response_metadata":{"usage":{"prompt_tokens":45,"completion_tokens":16,"total_tokens":61,"prompt_tokens_details":{"cached_tokens":0},"completion_tokens_details":{"reasoning_tokens":0}}},"tool_call_chunks":[],"id":"chatcmpl-AI4IhFB8OunsU857LKt9CtulAk1D5","usage_metadata":{"input_tokens":45,"output_tokens":16,"total_tokens":61,"input_token_details":{"cache_read":0},"output_token_details":{"reasoning":0}},"tool_calls":[],"invalid_tool_calls":[]}}],"agent":"agent"},"pending_sends":[],"metadata":{"source":"loop","writes":{"agent":{"messages":[{"lc":1,"type":"constructor","id":["langchain_core","messages","AIMessageChunk"],"kwargs":{"content":"It looks like you're testing the system. How can I assist you today?","additional_kwargs":{},"response_metadata":{"usage":{"prompt_tokens":45,"completion_tokens":16,"total_tokens":61,"prompt_tokens_details":{"cached_tokens":0},"completion_tokens_details":{"reasoning_tokens":0}}},"tool_call_chunks":[],"id":"chatcmpl-AI4IhFB8OunsU857LKt9CtulAk1D5","usage_metadata":{"input_tokens":45,"output_tokens":16,"total_tokens":61,"input_token_details":{"cache_read":0},"output_token_details":{"reasoning":0}},"tool_calls":[],"invalid_tool_calls":[]}}]}},"step":1,"parents":{}},"blobs":[["7ca17ee7-a17e-4ef5-ac72-4a8612daec9a","","messages","3","json",{"0":91,"1":123,"2":34,"3":108,"4":99,"5":34,"6":58,"7":49,"8":44,"9":34,"10":116,"11":121,"12":112,"13":101,"14":34,"15":58,"16":34,"17":99,"18":111,"19":110,"20":115,"21":116,"22":114,"23":117,"24":99,"25":116,"26":111,"27":114,"28":34,"29":44,"30":34,"31":105,"32":100,"33":34,"34":58,"35":91,"36":34,"37":108,"38":97,"39":110,"40":103,"41":99,"42":104,"43":97,"44":105,"45":110,"46":95,"47":99,"48":111,"49":114,"50":101,"51" etc....}],["7ca17ee7-a17e-4ef5-ac72-4a8612daec9a","","agent","3","json",{"0":34,"1":97,"2":103,"3":101,"4":110,"5":116,"6":34}]],"parent_checkpoint_id":"1ef89ccc-1d9b-69e0-8000-1520a807b509","pending_writes":[]}

[dosubot[bot]]

1. Customizing the Function: Instead of editing the function directly in node_modules, you can create a custom function in your project to override the existing logic. Define your own version of coerceMessageLikeToMessage and use it in your application. This way, you avoid modifying third-party code and maintain easier updates.

2. Redis Structure: The deserialization issue might be due to the structure of your Redis data. Ensure that the data stored in Redis matches the expected format for deserialization. The message objects should have clearly defined fields like type and content. If the data is stored as arrays, it might not be recognized as objects, leading to deserialization issues. Adjust the structure to ensure compatibility with the deserialization logic ^[1][2].

To continue talking to Dosu, mention @dosu.