CSV, Excel and Google Sheets in Famulor Knowledge Bases

Product codes, service descriptions, and reference prices often live in spreadsheets. Using them with an AI assistant has meant converting rows into prose or building an API for current values. Since September 5, 2026, Famulor can process CSV, TSV, and XLSX files, as well as Google Sheets connected through Drive, as structured knowledge-base sources. The feature supports exact identifiers, description-based search, and defined filters and aggregations (Famulor changelog, September 5, 2026).

The critical qualifier is that an imported table remains a snapshot from the last import, not a live system. Inventory, personalized balances, order status, and binding current prices still need a tool or API.

Key takeaways

• Famulor supports rectangular CSV, TSV, and XLSX tables, plus Google Sheets through a Drive connection.
• Table knowledge is a good fit for stable catalogs, service matrices, and reference prices with a visible effective date.
• Exact values, rows, counts, sums, minimums, and maximums are available within documented query boundaries.
• Google Sheets are not queried as a live database; the last import controls freshness, and formulas are not recalculated.
• Live, personalized, or writeback workflows still belong behind a mid-call tool or API.

What changed with structured table knowledge

The new capability refines an important architecture rule. General knowledge bases have traditionally served unstructured material such as manuals, policies, and FAQs. Treating a price table like a document made reliable cell-level retrieval difficult. Famulor’s table model instead preserves original cell values and column positions, along with source file, sheet, and row references (knowledge-base documentation, checked September 10, 2026).

An assistant can search for a complete product code, locate a row from its description, or use defined table queries. Results include rows, count, sum, minimum, and maximum under constrained filters. Joins, arbitrary expressions, currency conversion, and formula evaluation are unsupported.

The previous guidance in Knowledge Base vs System Prompt for AI Voice Agents therefore remains directionally sound but needs a post-update qualification: structured imported tables now support defined retrieval and aggregation tasks, while live data, personalized values, and writeback still require a tool or API.

When a spreadsheet knowledge base is the right fit

A table is a strong candidate when its rows describe stable reference objects and snapshot freshness is acceptable. Examples include a spare-parts catalog, a service matrix, a list of branch capabilities, a tariff table with an effective date, or an approved reference price list.

Ask four questions before importing:

1. Is the data stable enough for a snapshot? If a value changes minute by minute, a knowledge base is the wrong source.
2. Does each row represent a well-defined object? A product, service, or region should be described through consistent columns.
3. Are filters and simple aggregations sufficient? Multi-table logic, arbitrary calculations, and currency conversion need another system.
4. Should every imported sheet be indexed? Hidden XLSX sheets and every exported Google Sheets tab are included.

This separates knowledge retrieval from transactions. A catalog identifier may come from an approved import. “Is part X in stock right now?” requires a current system query, and a binding customer-specific quote should not be inferred from a general reference table.

Prepare CSV, TSV, XLSX, and Google Sheets before import

Source quality starts with a clean structure. According to the current Famulor documentation, column headings must appear in the first populated row. The table needs to be rectangular; merged cells and values beyond the width defined by the header are rejected. Put separate tables on separate sheets.

The documented format rules are specific:

• CSV can use commas or semicolons as delimiters; multiline values must be quoted correctly.
• TSV uses tab separators.
• Supported text encodings are UTF-8 or UTF-16 with a BOM.
• Legacy .xls files need to be saved as XLSX or CSV first.
• XLSX imports include every sheet, including hidden sheets. Google Sheets imports include every exported sheet.
• Formulas are not recalculated during import. Cached results can be marked unverified; missing results and cell errors remain unavailable.

Remove helper sheets, confidential notes, and unnecessary columns before import. Add valid_from or data_as_of for time-sensitive references so a conversation can state the source date.

Set up the table in Famulor in five documented steps

This flow follows the current Famulor documentation linked above.

1. Create the knowledge base

Open Knowledge bases → New, assign a clear name, and describe what the source contains. “DACH service catalog — September 2026 snapshot” communicates purpose and freshness better than “new price list.”

2. Add the table source

Upload a CSV, TSV, or XLSX file. For Google Sheets, first create a Drive OAuth connection under Automations → Connections, then select that connection when adding the Drive source. The documentation describes cloud-drive sync as Beta and a separate plan feature. It requires Beta Features under Settings → Workspace and follows the workspace’s current sync rates (Famulor documentation, checked September 10, 2026). Check the conditions in your own workspace rather than assuming universal availability.

3. Check processing status

Wait until the item reports that it is ready. If processing fails, inspect the format, file size, header row, merged cells, and sheet layout. A visible filename alone does not prove that a new searchable index was created.

4. Test with RAG Search

Use RAG Search with realistic questions. As an implementation practice, test an existing exact code, a nonexistent similar code, a description query, several matches, and a filter or aggregation. This is a recommendation, not a product guarantee.

5. Assign the knowledge base to the assistant

Select the knowledge base in the assistant editor. Famulor says that connection alone does not require a prompt change. Add workflow instructions for ambiguity, stale snapshots, or critical identifiers when needed.

What the assistant can ask of the table

For an exact identifier, search for the complete value in quotation marks. A similar-looking code is not evidence that the requested code exists. If several rows match, the assistant should clarify instead of choosing. Search results retain source references and positional column keys, including file, sheet, and row for traceability. That does not mean those references are automatically spoken to a caller.

Advanced implementations can use table queries through POST /knowledge-bases/{id}/search and the search_knowledge_base MCP tool. Under the current query documentation linked above, a query can return rows, count, sum, min, or max. It can combine up to five equality or range filters with AND; number and date ranges require an explicit type.

There is a firm trust boundary around numeric outcomes. Totals are rejected when matching rows contain missing, ambiguous, formula-based, or mixed-currency values. Do not calculate a full-table total by manually adding a handful of search excerpts. Use a table query—or a responsible backend when the operation is outside the supported set.

Practical example: a hypothetical B2B parts and service catalog

Suppose an industrial service company maintains an approved reference catalog. The miniature table below is deliberately fictional and contains no real customer or product data:

The source file contains these three example records:

• FP-2048-A: Standard filter package for MX-20, DACH region, EUR 189.00 reference price, effective September 1, 2026.
• SV-3100-D: Basic remote service for MX-30, DACH region, EUR 420.00 reference price, effective September 1, 2026.
• SV-3190-D: Plus remote service for MX-30, DACH region, EUR 690.00 reference price, effective September 1, 2026.

A caller might ask for “FP-2048-A.” Exact matching should identify that record; “FP-2048-B” must not be treated as present merely because it looks similar. A second question might be, “How many DACH service packages for the MX-30 have a list price below €500?” Equality and range filters plus count are designed for that request. The words list price and the effective date still matter in the response; this record must not silently become a binding current offer.

For voice channels, add an explicit confirmation routine as an implementation recommendation: have the assistant repeat critical codes in understandable groups and ask the caller to confirm before triggering a downstream step. This is not a claim that Famulor automatically spells or validates identifiers. Test likely pronunciations, background noise, and easily confused character sequences in real test calls.

Snapshot or live system? Make the boundary explicit

Knowledge-base data reflects the last import. Google Sheets are therefore not queried live for every question. Famulor does not recalculate formulas and does not support joins, arbitrary expressions, or currency conversion. These are architecture-defining constraints, not footnotes (Famulor documentation, checked September 10, 2026).

The separate Google Sheets integration and Microsoft Excel 365 integration pages cover workflow actions, triggers, and writeback. This guide deliberately stays with retrieval and constrained aggregation over imported table snapshots.

If a replacement or retry fails, the previous searchable index remains available and the error is reported, according to Famulor’s documentation. This preserves continuity, but it specifically does not prove that the new data was accepted. Check the status, snapshot date, and test set after every import.

Current limits at a glance

The values below come from the Famulor documentation linked above, checked on September 10, 2026, and may change over time:

Formal compliance with a size limit does not make a file semantically safe: ambiguous types, mixed currencies, or hidden tabs can still make a query unsuitable.

Privacy and telephone-operation checklist

According to the documentation linked above, Famulor retains the original table file and XLSX imports include hidden sheets. A practical governance rule follows: import only fields and sheets that are authorized and necessary for this purpose. This is implementation advice, not a compliance guarantee.

Before rollout, verify the following:

• Does the source contain personal customer or employee data that does not belong in a general catalog?
• Are hidden sheets, comments, and helper columns genuinely intended for the assistant?
• Is the Drive connection restricted to the required source and internally authorized?
• Does the table include an effective or snapshot date for prices and tariffs?
• Does the conversation avoid presenting reference values as current, binding promises?
• Is there a clarification, tool call, or human handoff for unconfirmed codes, multiple matches, and missing live data?

Telephone conversations add transmission risk: letters, digits, and similar-sounding identifiers can be misheard or misrecognized. Confirm critical identifiers and connect binding actions to a current system response. The assistant should state uncertainty rather than infer an answer from a similar match.

Conclusion: use table knowledge deliberately as a snapshot

CSV, TSV, XLSX, and Google Sheets can now be more than flattened documents in Famulor. As structured knowledge sources, they support exact identifiers, description search, and tightly bounded table queries. That fills a practical gap for product catalogs, service matrices, and approved reference price lists.

The useful design question is not simply “spreadsheet or API?” but “snapshot or live process?” If the source is stable, rectangular, and clearly dated, a table knowledge base is a strong candidate. If the workflow is current, personalized, or write-enabled, an authorized tool or backend remains responsible. Prepare the source carefully, test exact and ambiguous cases with RAG Search, and verify that each new import is actually ready. That keeps the feature within its documented strengths without turning reference knowledge into an accidental real-time promise.