Aeion BI + AI LLM Providers
Aeion BI's two AI surfaces — natural-language report builder ("show me revenue by category") and a plain-English narrative summary of a report's results — work with any LLM provider exposed via the Aeion AI gateway. OpenAI for general-purpose, Anthropic for complex reasoning, Azure OpenAI for BAA-required workloads, Google for GCP-native shops, local Ollama for sovereignty / cost. Per-tenant BYOK, per-tenant cost tracking, per-tenant compliance posture.
Provider Comparison for BI
`
Natural-Language Report Building
A user types: "Show me revenue by product category for the last 30 days."
Plain-English Result Narration
A user runs a report — say, churn rate by customer tier for the last two quarters.
Per-Tenant Provider Configuration
Each tenant configures its own AI behavior for BI, independent of every other tenant:
FAQ
OpenAI for general use and fast iteration. Anthropic for complex reasoning and large context. Azure OpenAI when a BAA is required. Ollama for data sovereignty or cost-sensitive workloads.
Yes. Every tenant configures its own key, so your cost and compliance posture are entirely your own.
Per-tenant, per-call, broken down by day, month, and feature.
Yes — logged with input and output, retained per your tenant's policy, and exportable to your SIEM.
The narration prompt is built from the executed report's own summarized results (row counts, per-column stats, top rows) — the AI is asked to describe those numbers, not to speculate about causes outside what the report returned.
Configure a fallback provider that kicks in automatically on failure — or fall back to a local Ollama model, which is always available regardless of any external provider's status.
Yes — pick a cheaper, faster model for simple queries and a premium model for complex reasoning, per feature.
Hardware-dependent — a modern consumer GPU running a 70B model typically returns in 5-15 seconds. Slower than a commercial API, but free per query and fully private.
Yes — per-tenant prompt template overrides, instruction tuning, and few-shot examples are all configurable.
Yes — each conversation thread retains its own context and history.
Pre-aggregation reduces how much raw data has to enter the prompt, results are cached where appropriate, and model selection can be tuned per query to balance cost and quality.