Answered

Code Interpreter can't access a Knowledge-source file for full-dataset calculations

(0) Share

Report

Posted on by CU28081217-4

Hi all,

I'm building a Copilot Studio agent that answers KPI questions (sums, counts, averages) over a manufacturing dataset. I want the agent to use Code Interpreter to compute accurate totals across my full dataset.

My setup:

3 files added as Knowledge sources (Type shows as "Files"):
1. AMTC_Insights_Data – the main data table (one row per work order)
2. Column_Source_Mapping – field/synonym dictionary
3. Business_Metric_Glossary – metric calculation rules

Generative orchestration is On

Code Interpreter is enabled

Agent instructions tell it to use Code Interpreter over the Insights data for any SUM/COUNT/AVERAGE

The problem:
When I ask for a total (e.g., "total liquidated hours for June 2026"), the agent correctly resolves the column, aggregation, and period — but it only computes over a handful of retrieved rows, not the full dataset. It explicitly responds with something like:

"The CodeTool requires the actual dataset file to be attached as an input file (a f_<guid> reference), which isn't available from the knowledge source search results. The search tool returns only summarized text excerpts, not the complete structured file."

So it seems Code Interpreter cannot read a Knowledge-source file as a compute input — Knowledge appears to be semantic search (returns text snippets), while Code Interpreter needs the file attached to the conversation as an actual input file. The two pipelines don't seem connected.

My questions:

Is this expected behavior — that Code Interpreter cannot load a Knowledge-source file for full-table computation?

What is the recommended pattern to run accurate aggregations over a full dataset in Copilot Studio?
- Attaching the file at runtime in chat?
- Passing the file via Power Automate as a file input to Code Interpreter?
- Moving to a tabular/database connector (Dataverse/SQL) and querying instead?

If Knowledge + Code Interpreter is not the right combination for full-file math, what's the best-practice architecture?

I've confirmed all three files are indexed as "Files" (not connectors), and the data file has finished indexing. The governance/mapping logic works perfectly — the only blocker is getting the complete dataset into the compute step.

Any guidance or confirmation would be hugely appreciated. Thanks! 🙏

Tags: Copilot Studio, Code Interpreter, Knowledge Sources, Power Platform

Categories:

Calling actions from Copilot Studio

Copilot Studio pre-built agents/templates

Copilot Studio skills development

I have the same question (0)

All responses (2)

Answers (1)

Sort by

Verified answer

Mohsin Ali 680 on at

Like
a
(0)

Report
Copy link

Link copied!
Hello @CU28081217-4 - Your understanding is mostly correct for the way the files are currently configured. An uploaded file configured as an agent Knowledge source is primarily used for grounding/retrieval, so you shouldn't assume that Code Interpreter will always receive the complete underlying file when the knowledge search returns relevant content.

However, Microsoft does support Code Interpreter analysis of structured data through two documented patterns:

An end user uploads the CSV/Excel file directly during the conversation.

A structured file is stored in a SharePoint Documents library configured as a knowledge source.

Now for your case, I would suggest moving `AMTC_Insights_Data` to a SharePoint document library and adding that SharePoint location as the knowledge source. With Code Interpreter and Work IQ enabled, Copilot Studio can use Code Interpreter against structured files from the SharePoint knowledge source for calculations such as aggregations.

https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-code-interpreter-structured-data

However, if your dataset is a persistent manufacturing dataset and you require deterministic KPIs such as SUM, COUNT, AVG and date/group filtering, I would suggest treating it as structured business data rather than knowledge content.

In that case, storing the records in Dataverse/Azure SQL and exposing an appropriate query/action to the agent would provide a more reliable architecture. Let the database/query layer calculate the KPI over the complete dataset and let the agent handle the natural-language interpretation and presentation of the result.

Additionally, please refer below the best practices and the limitations of using code-interpretor.

https://learn.microsoft.com/en-us/microsoft-copilot-studio/code-interpreter-for-prompts#best-practices-for-code-interpreter-in-prompts

https://learn.microsoft.com/en-us/microsoft-copilot-studio/code-interpreter-for-prompts#limitations

Was this reply helpful? Yes No
Mohsin Ali 680 on at

Like
a
(0)

Report
Copy link

Link copied!

Hello @CU28081217-4 - Just wanted to follow-up if the suggestions shared above really help to sort out the issue or pointed in right direction.

Was this reply helpful? Yes No