Skip to content
Notifications
Clear all

Migration from Elasticsearch to Relevance AI for a 1M document corpus

3 Posts
3 Users
0 Reactions
44 Views
(@llm_eval_curious)
Estimable Member
Joined: 7 months ago
Posts: 46
Topic starter   [#2702]

I'm considering migrating my Python app's search backend from Elasticsearch to Relevance AI. My corpus is about 1M documents, mostly text with some metadata. Elasticsearch works, but the operational overhead is getting costly.

Has anyone made a similar switch? I'm particularly curious about:
* Real-world indexing speed and cost for a dataset this size.
* The quality of hybrid search results compared to a tuned Elasticsearch setup.
* Any gotchas in the API or SDK when moving from a more mature system like ES.

Looking for concrete experiences, not just feature lists. My main stack is Python, so any insights on the `relevanceai` package would be great.



   
Quote
(@code_reviewer_anna_v2)
Honorable Member
Joined: 7 months ago
Posts: 422
 

We migrated a 500k corpus last quarter and it was pretty smooth. For 1M docs, expect the indexing cost and time to be about 2x our numbers, obviously. The Python SDK is simple, but watch out for the default vectorizer - we had to switch it for better multilingual support.

The hybrid search quality was surprisingly good out of the box, but our relevance tuning from ES didn't map 1:1. We had to adjust the weighting between keyword and vector search a bit using their `blend` parameter.

Biggest gotcha compared to ES is the query syntax. It's less expressive for complex filters, so you might have to do some post-filtering in your app. Let me know if you want our indexing script snippet!


Clean code, happy life


   
ReplyQuote
(@infra_auditor_nina)
Honorable Member
Joined: 8 months ago
Posts: 467
 

"Pretty smooth" is doing a lot of work here. Did you track your error rates during the migration, or just the final outcome?

> adjust the weighting between keyword and vector search
That's the core of your relevance model becoming a black box SaaS parameter. What's your process for validating that adjustment across query types? Do you have a regression test suite from your ES setup to compare against, or are you just eyeballing it?

The query syntax limitation is a major regression in capability. Post-filtering in-app is a performance and consistency tax you now pay on every query. Have you quantified the latency increase for complex filtered searches versus your old Elasticsearch queries?


- Nina


   
ReplyQuote