CSVAgent only seeing a small section of the data

[greenmotion]

I'm experimenting with the CSVAgent and no matter what I do, it appears to only see a small section of the data. For instance, when I pass in a CSV with 100 rows (each row being an application) and ask it how many applications there are, it tells me 5.

I'm using Ollama for the language models (tried llama3.1, mistral etc) and have tried changing the token size, but nothing seems to make a difference.

I'm sure this is a small silly thing, but I can't seem to figure it out. What am I missing?

Thank you.

[dosubot[bot]]

Langflow's CSVAgent doesn't impose any hardcoded limit on the number of rows it processes from your CSV file—the component simply passes your data to the upstream agent logic, which is designed to handle the full dataset unless otherwise configured. The most likely reason you're only seeing a small subset (like 5 rows) is that the underlying agent (from langchain_experimental) summarizes or samples the data before sending it to the language model, to avoid exceeding token/context limits—a common approach for LLM-based tabular agents. This means the model may only "see" a portion of your data when answering questions, regardless of token size settings in Ollama or the model you choose.

If you want to experiment, you can try adjusting the pandas_kwargs parameter in the CSVAgent to control how the CSV is loaded, but unless you have set any filtering options there, it's unlikely to be the cause. Also, double-check your CSV for formatting issues (like extra headers or blank lines), but based on the code, formatting problems would usually cause errors, not silent truncation.

If you need the agent to process all 100 rows, you may need to look into customizing the upstream agent logic or breaking your data into smaller chunks for analysis. There are no recent changes or issues in Langflow that would cause this behavior—it's a design limitation of how LLM agents handle large tables today.
Source 1
Source 2

To reply, just mention @dosu.

---

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

[greenmotion]

I'm not using an upstream agent. Instead I am using the CSVAgent response output, linked to a Chat output.