Your list is exactly the starting point for any serious evaluation. The "real cost at scale" vector is often miscalculated because teams model for query volume, but neglect the computational expense of full re-embeddings during regulatory updates. For SEC filings, that's not a variable cost, it's a scheduled, massive spend.
On the integration point: the "weird latency" you mention is frequently traceable to the serialization layer in the database client, not the network or the core database. I've benchmarked the LlamaIndex-Pinecone connector and observed a consistent 40-60ms overhead per call in a controlled environment before any data leaves the region. That overhead becomes a fixed tax on your p99 latency, which directly impacts user perception in a live app.
Regarding data isolation, the multi-tenant concern is valid, but the more common compliance failure I've seen is in secondary systems. A vendor might offer dedicated pods for your vectors, but their logging aggregation, monitoring dashboards, or even customer support tooling might inadvertently expose tenant metadata. You have to audit the entire data flow, not just the primary storage API.
—chris
You've set the correct evaluation criteria. I'd argue your second point, operational consistency, is undersold. "Eventually consistent" in this context isn't just about different answers; it can create a compliance audit trail that shows logically contradictory states if a query and its verification run seconds apart. For finance, you need linearizable reads, which rules out a surprising number of popular vector stores at their default configuration.
On integration pain, the latency isn't just weird, it's often non-deterministic. That config hell usually involves tuning timeouts and retry logic buried in the library's HTTP client. When the LlamaIndex connector abstracts this, you lose visibility into whether a timeout is from network saturation, database load, or just the client library's serialization struggling with a batch of unusually large vectors. This opacity is a risk vector itself.
The cost model for re-embedding is the true iceberg. Most SaaS pricing is optimized for query throughput, not for bulk re-indexing of your entire corpus after an embedding model update. You'll find the bill for that one-time full refresh can exceed a year of query operations.
Measure twice, cut once.