The results can be good, but there's room for improvement.

One of my go-to LLM tests is to provide a zero-shot prompt for ObjectScript code that calculates the distance between two points. As seen in this example (https://community.intersystems.com/ask-dc-ai?question_id=238163), which has some issues.

Most LLMs get close to answering this question, but they often fail with operator precedence and / or the square root function, either misspelling $ZSQR as $SQRT or hallucinating a math library that doesnt even exist, such as ##class(%SYSTEM.Math).Sqrt().

The problem stems from the token volume of other languages outweighing ObjectScript in the gradient descent during training. This causes solutions and functions from other languages to bleed into the ObjectScript code.

RAG powered by IRIS vector search is a good approach to address this, but it looks like it can be improved.

Generally, the LLM often has the right idea of how to answer the question, but not always with the correct details. To use an analogy, the LLM already has millions of recipes, but these might not match the ingredients in the kitchen. If we can tell it what ingredients to use, it will do a far better job at morphing a relevant recipe.

One strategy is to first ask the LLM to break down a question into smaller questions, such as "How do you calculate the square root of a number?" However, the DC AI still struggles with these simple atomic questions about the ObjectScript language:

https://community.intersystems.com/ask-dc-ai?question_id=239065
https://community.intersystems.com/ask-dc-ai?question_id=239114
https://community.intersystems.com/ask-dc-ai?question_id=239125

At a minimum, the vector database should also include these atomic facts about IRIS and ObjectScript.

One approach I've experimented with is to produce a compact document of atomic facts and either include the entire text in the prompt or have the LLM select which facts it thinks it needs first from a keyword list.

This results in a prompt reengineered by the LLM itself:

"InterSystems IRIS ObjectScript Reference: ObjectScript has strict left-to-right operator precedence, use brackets to ensure correct operator precedence, $ZSQR(num) is used to calculate the square root of the given number, $ZPOWER(num,exponent) is used to calculate the value of a number raised to a specified power. Task: How do you calculate the distance between two points."

With this approach, we see that the DC AI gives a much better response:

https://community.intersystems.com/ask-dc-ai?question_id=239152

I'm enthusiastic about this technology's potential. If you need a beta tester for future versions, please reach out. I'm also happy to contribute my ideas further if its of use.

Hi Scott,

Probably best to avoid modifying the Ens.MessageHeader properties, doing so might affect trace logging and potentially lead to unexpected side effects.

Here are a few alternative ideas....

1. Modify the MSH Segment: In the normalization process, tweak the sender or receiver field in the MSH segment to include a unique identifier that corresponds to the source service name. Then use the MSH to route the message.  
2. Use a Utility Method: Develop a utility method within a class that inherits from Ens.Util.FunctionSet. This method would read the source config name from the first message header in the session. You can then use this method in your router logic as it will be automagically included.  
3. Separate Normalization Processes: A config only option would be to create a normalization process for each service and then use that process name in the router logic.

Hey Paul,

Half agree if the OP requirements turn out to be a file to file use case.

If not, I wanted to defend the record mapper solution a little so as not to put off other readers from this approach.

As an example, I recently implemented a solution that had a million line CSV file. This would generate around 300,000 messages from a complex mapping. Each record was individually passed into a DTL which in turn produced a master file update message. These were then pushed into the local EPR (as per the EPR's requirements).

Yes, many messages, but not what I would call an overhead when taken into context of the frequency and need for M16 messages.

Bottom line, the solution I selected was the most maintainable solution. Almost zero glue code to go wrong and all maintanence managed inside DTL's. This is exactly what the record mapper was designed for.