MCP tool generates unreliable output for complex formulas

Bugs & Issues
1

The formula creating MCP tool passes instructions to Fibery AI to generate the formula but it misinterprets or drops specific instructions. Side issue, then the “Description” field is very long.

The suggestion is a “literal override” which could have a “verification” step in it so that the AI agents can run their own own loop of generating, verifying, and then updating with a verified literal formula.

Warning, the below is generated by AI but it’s actually pretty good context for what I ran into :slight_smile:

update_formula_field (and create_formula_field) MCP tool generates unreliable output for non-trivial formulas

The MCP tool takes a natural-language description and uses an AI generator to produce the formula. Even when the description includes the literal expression verbatim with explicit “do not modify” instructions, the generator silently rewrites it, often breaking the formula. Observed failures during a single session:

  1. Stripped a critical filter clause. Description contained [Settled Periods].Filter(Name = [This Financials].[Period Name]).Count() > 0 with explicit “must be preserved exactly”. The generator returned the formula with that entire clause removed, reverting to a simpler shape that produced wrong totals.

  2. Added space-prefixed namespace to [This <DB>]. Literal [This Financials] was rewritten to [This Budget/Financials]. The latter is invalid syntax and the saved formula errored on every row with “Reference to undefined field This Budget/Financials”.

  3. Stripped [This <DB>] entirely. When given a stronger “do NOT add the prefix” instruction, the generator over-corrected and replaced [This Financials].[Period Name] with just [Period Name], which then resolved to a non-existent field on the inner-filter iteration scope.

  4. Casing/bracket normalisation differs from UI parser. Generator output sometimes uses AND/OR while the UI parser only accepts and/or; or wraps single-word identifiers (Periods, Assessments) in brackets that the UI rejects.

Suggested fix: add a “literal mode” parameter (or detect when the description contains an explicit formula expression in backticks) that bypasses the AI generator and saves the supplied expression verbatim. The current behaviour effectively makes the tool unusable for any formula more complex than a single function call — for everything else, users have to fall back to pasting in the UI.