When using Solr as indexing backend, the tokenization of the string is performed using the StandardTokenizerFactory which treats punctuations and special characters as delimiters which results in the more indexed terms being associated with the associated vertex (document)
Also there's a LowercaseFilterFactory which makes lookup case insensitive.
This schema design doesn't work well for the current basic search enhancement (ATLAS-1880) causing a lot of false positives/negatives when querying the index.
The workaround/hack for this is to do an in-memory filtering when such schema violations are found or push the entire attribute query down to the graph which might be in-efficient and memory intensive. (Current JIRA will track this)
Correct solution would be to re-index the existing data with a schema change and not use the mentioned code workarounds for better performance of the search. (Should be taken up in separate JIRA)