Add CSV Tables to the Sparkforge AI Knowledge Base

By Sparkforge Support Team ·Last reviewed May 3, 2026

Table Search adds a CSV source type to the Sparkforge AI Knowledge Base. Upload up to 50,000 rows, and AI Agents answer plain-English questions on the data using semantic similarity, not keyword matching.

Table Search adds a CSV source type to the Sparkforge AI Knowledge Base. Upload structured data — customers, inventory, transactions, KPIs — and AI Agents answer plain-English questions about it during conversations using semantic similarity, not keyword matching.

What you get

Natural-language queries against table data — no formulas or filters needed.
Semantic search that returns context-aware answers from relevant rows and columns.
Large CSV support: up to 50,000 rows and 500 columns; pick the 20 most relevant for indexing.
Structured data coverage that web pages and documents miss.
Bot enablement across Chat, Voice, and Workflow AI agents.

How to add a table

Open AI Agents → Knowledge Base.
Edit an existing Knowledge Base, or click Create Knowledge Base and name it.
Click Add Source and choose Tables.
Upload a CSV via drag-and-drop or the file picker (max 50 MB).
Review detected columns and adjust data types if needed.
Click Done to index the table — progress shows chunking status.
Attach the Knowledge Base to your AI Agent (Chat, Voice, or Workflow AI).
Test with a sample question like Which customers have overdue invoices?

CSV requirements

Format: .csv only, UTF-8 recommended.
Size limits: 50,000 rows, 500 columns, up to 50 MB. Select the 20 most relevant columns for indexing.
Header row: first row must contain column names.
Schema detection: automatic, with an 80% confidence threshold.
Clean data: remove null values, hidden formulas, and merged cells.

If a CSV is rejected after format checks, the data itself likely has an issue. Open the file in a spreadsheet program or text editor to inspect.

How semantic search works here

Sparkforge converts each table row into vector embeddings, so the bot understands meaning rather than exact words. That powers queries like customers who complained about billing or orders shipped overnight last week. The engine compares the question to every row chunk and returns the most similar matches — no SQL.

Behind the scenes

Column types are detected automatically with 80% confidence.
Rows chunk into groups of five (max 2,000 characters) for efficient indexing.
Chunk metadata is stored so answers cite the correct records.

Frequently Asked Questions

Can I upload Excel files?

Not yet. Save or export your sheet as CSV first, then upload it as a Table source.

How long until a new CSV is queryable by my bot?

Usually a few minutes. The progress bar shows chunking and indexing status, and the source becomes available as soon as it finishes.

What are the size limits?

Up to 50,000 rows and 500 columns per location, 50 MB per file. Pick the 20 most relevant columns for indexing.

Do I need to write SQL?

No. The bot uses semantic similarity over row chunks, so plain-English questions like 'Which orders shipped overnight last week?' return matching rows directly.

Which AI agents can use a Table source?

Chat, Voice, and Workflow AI agents — attach the Knowledge Base containing the table to any of them.