Help: read value of a previous step in an LLM provider

[ashishs]

Self Checks

• I have searched for existing issues search for existing issues, including closed ones.
• I confirm that I am using English to submit this report (我已阅读并同意 Language Policy).
• [FOR CHINESE USERS] 请务必使用英文提交 Issue，否则会被关闭。谢谢！:)
• Please do not modify this template :) and fill in all the required fields.

1. Is this request related to a challenge you're experiencing? Tell me about your story.

We have implemented a custom Model Provider plugin which invokes a custom small language model.
We did this as our model generates long responses which requires streaming.
Our model requires dynamic value for a few input parameters, which are generated in previous steps of a chatflow.
These parameters change on every message.

Is there any method available in the Model Provider plugin to read the output of a previous step in the chatflow?

2. Additional context or comments

No response

[dosubot[bot]]

There isn't a direct method in the Model Provider plugin API to access outputs from previous steps in a chatflow. Dify's workflow engine handles parameter passing and only provides the plugin with the final, aggregated input for model execution. If your model needs dynamic values generated in earlier steps, you should configure your workflow or prompt templates to inject those outputs as part of the context or parameters passed to your plugin. This way, the required values are available when your provider is called. Workarounds must be implemented at the workflow or variable configuration level, not within the provider plugin itself. Standard Dify context passing and variable injection support this pattern if referenced correctly. For more on accessing raw LLM responses, see this discussion.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[ashishs]

How do I configure this to be passed as the context or parameters dynamically from the workflow? My LLM Model is invoked through the LLMNode.

[dosubot[bot]]

To pass outputs from previous workflow steps dynamically to your LLMNode, use Dify's variable system and prompt templating features:

• In the LLM node prompt editor, you can insert variables from upstream nodes by typing / or { to open the variable insertion menu. This lets you reference outputs from previous steps directly in your prompt as context or parameters. For example, if a knowledge retrieval node outputs result, you can associate it as a context variable in the LLM node and insert it into your prompt where needed. See docs

• If you need to combine or format multiple outputs, use a Template node with Jinja2 syntax to assemble them into a single string, then pass that string as context to the LLM node. See example

• Only string-type variables can be used directly in LLM node prompts. If your upstream output is an object or array, convert it to a string (e.g., using a Template node or code node) before passing it as context. See docs

• For more advanced logic, you can use Jinja2 templating in the prompt itself to manipulate variables dynamically.

• If you want to store intermediate results for use in later steps, use the Variable Assigner node to write outputs to conversation variables, then read and convert them to strings for the LLM node. See docs

Limitations: LLM nodes only support string-type variables; citation features are disabled if you use regular variables instead of context variables; model parameters (like model name, temperature) must be statically defined in the workflow and cannot be changed dynamically. See issue

If you run into issues with missing variables or conversion errors, double-check that your variable types match what the LLM node expects and that all referenced variables are available in the workflow's variable pool.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}