URL detection is overzealous

[CalebCourier]

Inconsistently, but often enough to notice, chunks of a sentence will be flagged as a URL when they are not.

Example:

from guardrails import Guard
from guardrails.hub import CompetitorCheck, DetectPII

guard = Guard().use(
    CompetitorCheck(
        competitors=["Fortran", "Ada", "Pascal"],
        on_fail="fix"
    )
).use(DetectPII(pii_entities="pii", on_fail="fix"))

response = guard.validate("The author is Paul Graham. Growing up, he worked on writing short stories and programming, starting with an early version of Fortran on an IBM 1401 in 9th grade. Later, he transitioned to microcomputers and began programming more extensively, including writing simple games and a word processor on a TRS-80.")

print("Raw: ", response.raw_llm_output)
print("Guarded: ", response.validated_output)

Raw:  The author is Paul Graham. Growing up, he worked on writing short stories and programming, starting with an early version of Fortran on an IBM 1401 in 9th grade. Later, he transitioned to microcomputers and began programming more extensively, including writing simple games and a word processor on a TRS-80.

Guarded:  The author is <PERSON><URL>owing up, he worked on writing short stories and programming, starting with an early version of [COMPETITOR] on an IBM 1401 in 9th <URL>er, he transitioned to microcomputers and began programming more extensively, including writing simple games and a word processor on a TRS-80.

I'm guessing it may have to do with how the chunking removes spaces so it sees .[some three letters] and thinks it's a top level domain. For example turning 9th grade. Later, into 9th <URL>er, kind of makes sense because .lat is a known top level domain. However, the other assumably flagged tld is .gro which isn't a thing unless it's fuzzy matching to .org.

[dtam]

the underlying url regex in presidio doesnt flag this might be coming from context or one of the nlp code paths in presidio