LLM on existing lakehouse data
Data already lives in Databricks or Snowflake; want governed inference/RAG without exporting corpora.
At a glance — Oct 2026
Updated 2026-10-27 · forward this page: /jtbds/lakehouse-llm#briefScores as of 2026-09-14 (43 days old; Depot marks scores stale after 30 days). The pick is still Depot's current editorial pick; it has not been re-reviewed since. Prices below carry their own dates.
- Depot Pick
- Databricks Mosaic AI · 7.11/10 for this job (curated estimate, medium confidence)
Data gravity + privacy/governance beat raw DX.
- Runner-up
- Snowflake Cortex · 7.10/10
Scores higher than the pick on: Price / performance, Developer & agent DX, Privacy / data retention.
- When not to use the pick
- Greenfield agent with no lakehouse — use direct APIs + pgvector/Qdrant.
- Lock-in / migration
- The pick has lower stored lock-in risk than runner-up Snowflake Cortex (7.8 vs 8).
- Sourced price (usd_per_million_input_tokens)
Pick: no sourced price in this job's unit (no_sourced_price).
Cheapest matching SKU: Amazon Bedrock Gemma 3 4B $0.04 — source, as of 2026-10-03, re-checked 2026-10-27. Not the pick; not quality-adjusted.
Price history: first seen 2026-10-03 · last changed no change recorded · history JSON
- Free option
- No free resource lists this job. A cheap paid SKU is not a free option.
- Failure modes to watch
- Runaway spend: Set a budget and a tool-call cap per task. Alert when spend jumps. Stop the run at the cap instead of summarizing a partial trace as success.
- Collateral side effects: Snapshot the workspace or app state, then diff it against an allowlist. A passing check plus an unexpected diff is a failure.
- Hallucinated tool arguments: Reject a call that fails the current schema or names an unknown id, and do that before the tool runs a side effect.
- Partial tool success: Require a per-item status and a failure count. Success for the task requires the failed count to be zero or explicitly handed to a retry queue.
- Schema drift: Validate every tool call and every loaded table against the current schema. Quarantine mismatches instead of coercing them.
- Silent truncation: Tools must return whether the payload was truncated and how long the original was. Fail the task if the answer depends on the missing span.
Cite: Agent Provider Depot, Depot Pick for lakehouse-llm: databricks-mosaic, scores as of 2026-09-14. Agents: GET /api/decide?jtbd=lakehouse-llm
All sourced prices, cheapest first · for data-etl agents
Databricks Mosaic AI
Data gravity + privacy/governance beat raw DX.
Pricing (directional): As of 2026-09-14: Lakehouse-native model training/serving billed via Databricks DBUs + cloud compute. Foundation Model Serving and Mosaic tooling sit inside Databricks contracts — not indie per-token pricing. Strong if data already lives in…
Lock-in risk: higher = more lock-in (worse). Overall already inverts this dimension.
Pick only when your data/platform already is Databricks. Otherwise Bedrock/Vertex or direct APIs.
Quote for LLM on existing lakehouse data
Depot Pick stays Databricks Mosaic AI. Prices below are sourced SKU rows only (as of 2026-10-27). Input and output are not blended. GET /api/quote?jtbd=lakehouse-llm
Pick is not a price winner. Cheapest sourced input $/1M among candidates is amazon-bedrock at 0.04. The pick has no priceSortEligible input row.
Sourced price unavailable for the Depot Pick. No row in the pricing index.
2 of 4 candidates have no sourced price row. 2 are excluded from the Input $/1M comparison (chat_completion_input). min sourced input SKU among candidates; not quality-adjusted
Runners-up
Snowflake Cortex
Use when Snowflake already holds the data. Don't drag agent backends into Snowflake just for Cortex.
Amazon Bedrock
Enterprise multi-model switchboard. Great when AWS is mandatory; mediocre when you just want the best Claude…
Google Vertex AI
Enterprise Gemini path. Use AI Studio for prototypes, Vertex when security review starts asking questions.
When NOT to use the top pick
Greenfield agent with no lakehouse — use direct APIs + pgvector/Qdrant.
Full ranking (score table)
Ranked with this JTBD's weights. Scores are curated estimates unless marked otherwise. Agents: prefer /api/decide then this table for tie-breaks.
| # | Provider | Overall | Fit | $/perf | Latency | DX | As of |
|---|---|---|---|---|---|---|---|
| 1 | Google Vertex AI | 8.0 | 9.0 | 7.0 | 9.0 | 7.2 | 2026-09-14 |
| 2 | Amazon Bedrock | 7.8 | 8.8 | 6.0 | 9.2 | 6.8 | 2026-09-14 |
| 3 | Databricks Mosaic AIDepot Pick | 7.1 | 8.4 | 4.8 | 8.5 | 6.8 | 2026-09-14 |
| 4 | Snowflake Cortex | 7.1 | 7.8 | 5.5 | 8.4 | 7.0 | 2026-09-14 |
Score methodology for this job
Data gravity + privacy/governance beat raw DX.
capabilityFit: 25%pricePerformance: 18%latencyReliability: 15%developerDx: 15%privacyPosture: 10%lockInRisk: 7%maturity: 10%