Should synonyms work with filtering?

[ctorgalson]

Description

We were looking at synonyms (specifically one-way synonyms) as a way to improve a client's search. The search uses filter_by in order to only return exact matches. But there's a special case of search where synonyms would be convenient.

Testing from the cli, I find that a one-way synonym works exactly as expected when using an ordinary search, but not at all when using filter_by for the search.

Am I seeing:

• [ ] the correct behaviour?
• [ ] a planned behaviour?
• [ ] a bug?

Steps to reproduce

Returns synonyms in results

http://typesense/collections/vehicles/documents/search
  ?q=volkswagen golf
  &query_by=vehicle
  &include_fields=make,model,year,price

Doesn't return any synonyms in results

http://typesense/collections/vehicles/documents/search
  ?q=*
  &query_by=vehicle
  &filter_by=vehicle:=[volkswagen golf]
  &include_fields=make,model,year,price

Expected Behavior

I don't know. I can't tell from the docs if what I'm trying is expected/unexpected/desired/undesired :grinning:

Actual Behavior

No synonyms in results when using a filter_by query.

Metadata

Typsense Version: 0.20.0

OS: any

[kishorenc]

@ctorgalson Synonyms are not used when resolving the filter query. I'm open to adding support for this, if there is a strong business case. Synonyms currently work only with query string because the query string is a user input, while filter_by queries are constructed by the application, so there is not much scope for ambiguity.

[ctorgalson]

Thanks @kishorenc. I guess we may not have a strong business case--especially since we probably have another approach to this, but here's the reasoning for why we looked at this as an option:

• The search is a make & model vehicle search,
• The client's basic search strategy is to provide autocomplete-style search suggestions with the idea that a search for "volkswagen golf" returns only Volkswagen Golfs (in the current application at least, this is simpler with a filter query),
• But there are variants of models like the "Golf SV" and "Golf GTI",
• In this case, the client's preferred behaviour is identical to a one-way synonym, namely:
  • a search for "volkswagen golf" would return all results matching "volkswagen golf *", and
  • a search for "volkswagen golf gti" would return only "volkswagen golf gti" results,

---

The alternate approach is to index the make + model as string[] like this:

['Volkswagen Golf', 'Volkswagen Golf SV',],
['Volkswagen Golf', 'Volkswagen Golf GTI',],
['Volkswagen Golf',],

That way, all three results exactly match volkswagen golf, but only one each matches volkswagen golf sv and volkswagen golf gti. This approach works well (we're using this strategy already in a related case), but it shifts the load onto the indexing routine since we need to know/check that an item has synonyms when that item is indexed (i.e. instead of maintaining a small list of synonyms separately).

But again, this isn't insurmountable.

[jsutaria]

+1