RecommendationFinder is used in several places to produce "Did you mean...?" suggestions when variables or keywords are not found. It performs fuzzy string matching over all candidates, which is O(n×m) per lookup.

Affected locations

1. robot.variables.notfound.variable_not_found() — iterates over all variable candidates
2. robot.running.namespace.KeywordRecommendationFinder — iterates over all keywords from all libraries and resources (also affects --dryrun, since KeywordRunner passes recommend_on_failure=True by default)

(A third location in parsing.lexer.settings operates on a small fixed set of setting names, so it is not a performance concern.)

Note that there is already a TODO comment in

For tools that use Robot Framework's variable resolution programmatically (e.g., IDE language servers, static analysis tools, like RobotCode ), variable_not_found() may be triggered thousands of times during analysis. In large projects with many variables, this becomes a critical bottleneck.

Real-world impact: Analyzing 375 Robot Framework files took over 30 minutes, with the vast majority of time spent in RecommendationFinder. After monkey-patching variable_not_found to skip recommendations, the same analysis completes in under a minute. See: robotcodedev/robotcode#587

The KeywordRecommendationFinder in namespace.py has the same O(n×m) characteristic, though it primarily affects runtime execution and --dryrun rather than static analysis. RobotCode does not use RF's runtime namespace for keyword resolution, so this is not triggered in our case, but other tools or --dryrun with many missing keywords could be affected.

Current workaround

In RobotCode, we now monkey-patch variable_not_found across all modules that import it (robot.variables.notfound, finders, evaluation, store, robot.variables) to replace it with a version that simply raises VariableError without calling RecommendationFinder. This works, but it's fragile — it depends on internal module structure and could break with any refactoring of the robot.variables package.

Suggestion

A simple global flag like RECOMMENDATIONS_ENABLED (e.g., on RecommendationFinder or in robot.utils.recommendations) that can be set to False to skip all fuzzy matching would be sufficient and avoid the need for monkey-patching.