How can I extract and format JSON from an HTTP API response for LLM analysis?

Below is an overview of some strategies you can use to extract data from your JSON response and prepare it for further analysis by an LLM:

---

1. Use JMESPath to Filter or Transform JSON

n8n supports the JMESPath query language, which lets you extract, transform, or reformat your JSON data before sending it to the LLM. For example, if your API response contains an array of log entries and you only want certain fields, you can do something like this in an Expression or Code node:

In an Expression (JavaScript):

JSCopied
{{$jmespath($json, "logs[*].message")}}

In a Code Node (JavaScript):

JSCopied
let logMessages = $jmespath($json, "logs[*].message");
return { logMessages };

This will extract all the messages from your log entries so that your LLM only sees the relevant parts of the data.

---

2. Use the Custom n8n Workflow Tool

Once you have extracted the data you need, you can pass this structured information as input to an LLM. The Custom n8n Workflow Tool lets you plug in workflows as tools for your LLM. You can create a sub-workflow that processes the JSON data further (perhaps filtering, aggregating, or reformatting it) and then have the LLM call this tool as needed.

For example, if you have already extracted your log messages, you could set up a sub-workflow that reformats them or even extracts statistics before returning a final prompt for the LLM.

---

3. Leverage AI Workflow Components

Depending on your needs, consider combining multiple nodes:

• Chat Trigger: To start your workflow on demand.
• Agent: The core node that interacts with other parts of your workflow and can decide to use your custom tool.
• LLM Chain Nodes: Use these nodes along with output parsers (like the Auto-fixing Output Parser or Structured Output Parser) to help your LLM understand the refined JSON data, ensuring that any unexpected formatting is corrected before analysis.

For example, your workflow might first extract the data with a Code node using JMESPath, then pass it to your LLM chain:

JSCopied
// Extract the 'message' field from each log entry.
let messages = $jmespath($json, "logs[*].message");
return { messages };

You can then reference these messages in your prompt to the LLM node (e.g., setting a parameter like prompt: "Analyze these log messages: {{ $json.messages }}").

---

Summary

• Extract with JMESPath: Use n8n's built-in jmespath() method to selectively extract or transform the JSON response.
• Reformat Data: If needed, further format the extracted data using Code nodes or sub-workflows.
• Feed to LLM: Once the data is refined (and optionally structured via output parsers), pass it into your AI workflow using nodes like the Chat Trigger, Agent, and LLM Chain.

This approach keeps your workflow modular and ensures that the LLM gets only the data it needs in a clear, structured format for analysis.

Below is a streamlined approach you might consider:

1. Use JMESPath with a Code Node

n8n supports the JMESPath library and a custom method called jmespath(). You can use this in a Code node or as an expression to extract just the relevant parts from your API response. For example, if you only need error logs from your system, you could use something like:

JSCopied
let errors = $jmespath($json, "logs[?level=='error']");
return { errors };

This will give you a JSON structure with only the error logs that you can then forward to your LLM.

2. Consolidate Data into a Single JSON Object

When working with LLMs, it’s often efficient to reduce latency by combining multiple steps into one request. Instead of sending raw JSON, format your data into a JSON object with clearly named fields. For example:

JSCopied
let errors = $jmespath($json, "logs[?level=='error']");
let warnings = $jmespath($json, "logs[?level=='warning']");
return {
  errors: errors,
  warnings: warnings
};

This way, you can easily reference specific information in your prompt for the LLM.

3. Provide Clear JSON Instructions to the LLM

When you pass the JSON to the LLM, ensure that your prompt instructs it to treat the data as JSON. This is especially important in JSON mode so that the model doesn’t generate extraneous whitespace or partial responses. For example, use a system or user message like:

Please analyze the following logs provided as JSON. Make sure your response is formatted as JSON.

Summary

• Extract: Use n8n’s jmespath() in a Code node or expression to pick out the logs you need.
• Consolidate: Combine those results into a structured JSON object.
• Instruct: Provide a clear prompt ensuring the model returns its output in JSON format.

This strategy not only cleans up your input but also makes it easier for the LLM to analyze and extract insights from the logs.