Everyone's rushing to slap "AI-powered search" onto everything, and the marketing blurbs for both Relevance AI and Marqo read like they were written by the same over-caffeinated growth hacker. "Revolutionize your user experience with intelligent, semantic, vector-powered blah blah." Spare me.
I'm evaluating for a concrete use case: ingesting video transcripts, generating metadata (topics, entities, sentiment), and letting users search through this combined dataset. Think: "find all clips where the speaker is positive about cloud costs but mentions vendor lock-in." It's not just a simple vector search; you need hybrid search combining keywords and vectors, and ideally some structured filtering on the extracted metadata.
Relevance AI pitches its "Workflows" and "AI Agents" hard, but that's just extra fluff I don't need. I want to know about the actual search engine. Marqo, being open-source, at least lets me see the guts. But their documentation spends more time on the "getting started" hype than on the nitty-gritty of complex queries.
My main concerns are vendor lock-in and cost predictability. Relevance AI is a SaaS black box. What's the real query cost at scale? Can I export my vectors and metadata cleanly if I need to run? Marqo lets me self-host, but then I'm managing the infrastructure. Their cloud offering is newer, so who knows where the pricing will go.
For a hybrid search on mixed data (text and metadata), the API design is critical. Here's a simplified example of the *type* of query structure I'd need, comparing a hypothetical approach:
```json
// What I need conceptually
{
"query_text": "positive sentiment on open source",
"filters": {
"metadata.entity": "self-hosting",
"metadata.topic": "cloud cost"
},
"hybrid": true,
"alpha": 0.5 // weight between lexical and vector search
}
```
Which platform actually lets me do this cleanly without jumping through five "agent" hoops or writing convoluted DSL? Does Marqo's Sparse-Dense hybrid search handle the structured filter as a first-class citizen, or is it an afterthought? Does Relevance AI's search API force me through their workflow builder for this?
Looking for real implementation experience, not a recitation of their feature lists.
Just my 2 cents
You've nailed the core issue: the marketing is all fluff until you hit the scaling cost and lock in. With your video metadata case, the structured filtering is the killer. Both can probably do hybrid search, but the moment you need complex boolean logic on those extracted topics and sentiments alongside vector similarity, you'll hit limitations.
Relevance AI's SaaS black box will kill your cost predictability. Query cost at scale is their secret sauce, and you won't know until the bill arrives. Export? Good luck. You're renting their brain. Marqo being open-source lets you at least run it yourself and see the query engine, but their distributed mode is young and operational overhead is real.
Frankly, for your specific "positive sentiment + vendor lock-in" query, you might be better served by a more mature search stack like OpenSearch or Elasticsearch with a built-in vector plugin. You'd own the ingestion pipeline for transcripts and metadata, which is the complex part anyway. Then you have full control over hybrid queries and filtering, no black box surprises.
Been there, migrated that