Bug: Agent Chat View Fails to Re-run Tool, Causing on consecutive Second Call

BUG REPORT for KNIME Analytics Platform Version: 5.7

When using a tool within the new Agent Chat View node, the tool executes successfully on its first run. However, if the same tool is called a second consecutive time, the agent fails with a

RuntimeError

indicating a file or directory was not found.

The error message always points to a temporary folder that no longer exists, suggesting the agent is incorrectly reusing a stale file path from the first run instead of re-executing the tool’s workflow.

Error Message:

RuntimeError: Tool execution failed with: Workflow contains one node with execution failure:
Vector Store Retriever #[node_id]: [Errno 2] No such file or directory: ‘/var/folders/r3/[...]/embeddings’

Definitive Proof of Caching Issue

I have definitively confirmed that the tool’s workflow is not being executed at all on the second run.

To test this, I placed a

CSV Writer

node at the start of my tool’s workflow, configured to append a unique “cache buster” variable (the current timestamp) to a file every time the tool runs.

*

On the first successful tool call, the CSV file is correctly updated with the timestamp.

*

On the second, failing tool call, the CSV file is not updated.

This proves the Agent Chat View is not re-entering and re-running the tool’s component. It is failing based on a cached state before the tool’s workflow can even begin.

Steps to Reproduce

The issue can be reproduced using a standard, best-practice RAG setup:

Workflow A (Indexing): Create a permanent vector store using the

FAISS Vector Store Creator

node, saving it to a fixed local directory (e.g.,

/data/my_vector_store

). Run this workflow once.

Workflow B (Tool Component): Create a component for the agent tool. This workflow reads the permanent vector store using a

FAISS Vector Store Reader

node and passes the data to a

Vector Store Retriever

node. The path to the permanent store is explicitly defined.

Agent Workflow: Use the

Agent Chat View

node and configure it to use the tool from Workflow B.

1.

Execution:

Start the agent and ask a question that triggers the tool. The first run will succeed.

1.

Immediately ask a second question that triggers the same tool. The agent will fail with the `[Errno 2]` error.

Troubleshooting Steps Already Taken

I have already attempted all standard solutions to prevent caching and state issues, none of which have resolved the problem because the tool workflow is not being re-executed. These include:

*

Ensuring the Vector Store is in a permanent, local directory.

*

Using a `FAISS Vector Store Reader` to load the store on each call.

*

Forcing the permanent file path into the reader/retriever nodes using Flow Variables.

*

Adding a “cache buster” flow variable (a unique timestamp) connected directly to the

Vector Store Retriever

node to try and force re-execution.

None of these methods work, as the agent’s caching prevents the tool’s workflow from running a second time.

This appears to be a bug in the state management or caching logic of the

Agent Chat View

node. Any guidance or a potential fix would be greatly appreciated.

Thank you!