Contribute advanced Hybrid search example in OpenAI Cookbook (python)

[tylerhutcherson]

The existing cookbook just touches the surface:
https://github.com/openai/openai-cookbook/blob/main/examples/vector_databases/redis/getting-started-with-redis-and-openai.ipynb

Contribute a Python notebook that demonstrates complex Hybrid queries with Redis VSS and other search features (an ecommerce dataset might work nicely) including

• Numeric range filters
• Tag filters
• Full text search "filters"
• Client-side hybrid scoring combing both BM25 lexical AND semantic search. This could be done in a pipeline to send 1 redis call to fetch both search results (top K) and then merge the sets. Show performance improvement with this technique over pure lexical or pure semantic?

[michaelskyuan]

Submitted PR

[tylerhutcherson]

Initial review submitted from our end >>> openai/openai-cookbook#417

@michaelskyuan at some point we will also want to make an update to this notebook that covers bullet point 4 above. This is a bit "green field" in the sense that we have not yet explicitly tried this. But it's theoretically possible to do true weighted hybrid search using a redis pipeline command and "merging" results from the two scoring algorithms (BM25 + KNN/CosineD). I sense the lift will be a bit more on this, and since not immediately pressing, I will spin it off into a separate issue that we can re-prioritize when the time is right, probably in the next month.

[michaelskyuan]

I agree @tylerhutcherson. And I believe this topic deserves it's own separate notebook with a denser text dataset.
Let's leave OOTB Redis Hybrid search functionality on the current notebook and have a specific notebook that will address normalization of lexical and semantic scoring using a more appropriate dataset instead of an ecommerce dataset.

[Spartee]

@michaelskyuan This App/notebook was recently contributed by OpenAI. We could use some of this.

[ladrians]

I am struggling trying to apply a hybrid filter similar to these examples, the difference I am using langchainjs 0.2.16 from typescript with the standard bindings, using RedisSearch 2.6 from the cloud.

I have the following metatadata associated to a chunk item

{\\\"name\\\"\\:\\\"20030331\\\",\\\"description\\\"\\:\\\"2003/03/31\\\",\\\"year\\\"\\:2003,\\\"month\\\"\\:3,\\\"day\\\"\\:31,\\\"date\\\"\\:20030331,\\\"id\\\"\\:\\\"5251cc41\\-307d\\-4117\\-9b0f\\-9e408eb37011\\\",\\\"doc_id\\\"\\:\\\"5251cc41\\-307d\\-4117\\-9b0f\\-9e408eb37011\\\"}

Would like to filter in a range for the year for example similar to the following:

@metadata:(\\\"year\\\"\\:[(2001 2004])

always returning 0 records. If I use exact match or the negaction the case works. I get started with

filterQuery = `@metadata:(-\\\"year\\\"\\:2001)`;
itemList = await client.ft.search(chunk, filterQuery);
console.log(filterQuery, itemList.total);

OK, 3 items returned

@metadata:(-\"year\"\:2001) 3

Then I started trying with a range

filterQuery = `@metadata:(\\\"year\\\"\\:\\[2003 2005\\])`;
itemList = await client.ft.search(chunk, filterQuery);
console.log(filterQuery, itemList.total);

should return elements buyt I get

@metadata:(\"year\"\:\[2003 2005\]) 0

Very similar to the previous one

filterQuery = `@metadata:(\\\"year\\\"\\:\\[\\(2003 2005\\])`;
itemList = await client.ft.search(chunk, filterQuery);
console.log(filterQuery, itemList.total);

should return elements too but still I get 0

@metadata:(\"year\"\:\[\(2003 2005\]) 0

Trying other filters and using And (&) worked fine:

filterQuery = `@metadata:(\\\"year\\\"\\:2003&\\\"date\\\"\\:20030331)`;
itemList = await client.ft.search(chunk, filterQuery);
console.log(filterQuery, itemList.total);

OK, 1 item filtered

@metadata:(\"year\"\:2003&\"date\"\:20030331) 1

Got some ideas from the referenced notebook but still cannot make it work. any idea?

thanks in advance!