Question 1

Where does inference actually run?

Accepted Answer

By default, on your own iron — local Ollama serving your chosen model (Qwen, Llama, Mistral, etc). In hybrid mode, you can route the LLM call to a cloud provider while keeping retrieval and audit fully local. In fully-cloud mode, both retrieval and chat go through your chosen OpenAI-compatible endpoint. The provider toggle lives in admin settings, takes effect within 30 seconds, and is itself audited.

Question 2

What does the audit trail actually contain?

Accepted Answer

For every interaction: timestamp, user ID, trace ID, query text, retrieval scores, source chunks (with PII redacted), generated answer, latency per stage, model used, guard decisions, and feedback. Rows live in Postgres for 90 days by default and can be exported as CSV. Enterprise deployments can archive to immutable/WORM object storage indefinitely.

Question 3

Does it work air-gapped?

Accepted Answer

Yes. The reference deployment is fully offline-capable. The container images, model weights, embedding models, and reranker can all be mirrored to your internal registry. No outbound calls are required at runtime.

Question 4

How is PII handled?

Accepted Answer

Three layers. The input guard detects and optionally blocks PII before it enters the pipeline. The content filter scrubs PII out of retrieved chunks before they're sent to the LLM. The output guard checks the final response for hallucinated PII patterns. You configure detection rules per workspace.

Question 5

Can we bring our own model?

Accepted Answer

Yes — anything with an OpenAI-compatible chat completion endpoint works. Same for embeddings (Nomic, BGE, your own fine-tune) and the optional reranker. Switch via admin settings; existing audits remain intact.

Question 6

How does grounding actually work?

Accepted Answer

Every answer is generated with the retrieved chunks as the only context. The system prompt requires the model to either cite a source or say 'I don't have enough information'. The output guard rejects responses without at least one citation. If the corpus is silent, the model is silent.

Question 7

What's the cost of a typical pilot?

Accepted Answer

€0 for up to 60 days, up to 5 seats and 10k queries. You provide the hardware (a single A100 or 2× L40S handles most pilots comfortably) and we provide the system, deployment support, and a dedicated Slack channel.

Question 8

Do you have SOC 2 / ISO 27001?

Accepted Answer

The product itself is shipped as software you run; certifications live with whoever operates the deployment. For Enterprise customers, we scope a controls mapping engagement (SOC 2, ISO 27001, NIST 800-53) that maps every audit field, guard, and span to the relevant control families.

Documentation