# Best way to handle very large tool outputs in LangGraph (avoid LLM + LangSmith overload)

**URL:** <https://forum.langchain.com/t/best-way-to-handle-very-large-tool-outputs-in-langgraph-avoid-llm-langsmith-overload/2445>\
**Category:** LangGraph\
**Tags:** intro-to-langgraph, python-help\
**Created:** [December 7, 2025, 11:18pm UTC](https://forum.langchain.com/t/best-way-to-handle-very-large-tool-outputs-in-langgraph-avoid-llm-langsmith-overload/2445 "2025-12-07T23:18:29Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![csarww](https://avatars.discourse-cdn.com/v4/letter/c/0ea827/32.png) [@csarww](https://forum.langchain.com/u/csarww)\
**Post date:** [December 7, 2025, 11:18pm UTC](https://forum.langchain.com/t/best-way-to-handle-very-large-tool-outputs-in-langgraph-avoid-llm-langsmith-overload/2445/1 "2025-12-07T23:18:29Z")

</div>

Hi all — I’m building an agentic workflow (LangGraph + Gemini) where one of my tools returns a **very large JSON output**. The full data is required for downstream processing, but:

- I **must not** send the raw JSON to the LLM

- I **must avoid** huge payloads appearing in LangSmith traces

- I’m running **local-only and [Streamlit.io](http://Streamlit.io)**, no external DB

- The agent may generate **multiple large artifacts** per run

I tested three approaches:

* * *

### **1. `response_format="content_and_artifact"`**

Store the large JSON in `artifact`, return only a summary in `content`.  
With a small patch, LangSmith won’t upload the full artifact.

**Pros:** agent-friendly, clean, scalable  
**Cons:** still need disk storage if I want persistence

* * *

### **2. Write large JSON to local temp files**

Tool writes to disk and returns only a file key/path.

**Pros:** safe, no risk of LLM exposure or LangSmith overload  
**Cons:** manual lifecycle, less native to LangGraph

* * *

### **3. Global variables**

Simple dict holding all results.

**Pros:** easy to prototype  
**Cons:** breaks in reruns, not persistent, not agent-friendly

* * *

# **Question:**

**What is the recommended pattern for handling large tool outputs that must remain accessible across an agent’s workflow while keeping them out of the LLM context and LangSmith traces?**

Should I primarily rely on `content_and_artifact` + disk-backed storage, or is there a more idiomatic approach?

Thanks!