Common questions
Frequently asked questions about using BullSequana AI.
Platform overview and access
What is BullSequana AI?
BullSequana AI is a modular platform for running AI and data workloads in production. It combines a Kubernetes-based runtime foundation with user-facing and developer-facing services for model access, data handling, retrieval, automation, observability, and deployment.
What can I do with it?
Depending on your role, you can:
- use the
Portal - access models through the stable
BSQAI API - upload and manage files
- build retrieval and document-based AI experiences
- deploy models and AI services
- run analytical and data-oriented use cases
How do I access the platform?
You access the platform through the URL provided for your deployment. Authentication is required and is normally handled through your organization’s identity provider via the platform login flow.
Do I need technical skills to use it?
Not always. Many day-to-day actions are available through the portal, especially for business users and operational users. More advanced work, such as API integrations, workflow design, or deploying new use cases, is meant for developers, AI engineers, or platform teams.
Data, files, and vector search
Where is my data stored?
Different types of data are stored in different platform components:
- S3-compatible object storage,
Rook CephRGW or an external S3 service, for files, objects, and artifacts. Each tenant can have its own bucket for its library files, sandbox workspaces, and MLflow artifacts. PostgreSQLfor application and operational dataMilvusfor vectorized data used in semantic retrievalLakekeeperwith Apache Iceberg for governed lakehouse tables
Can I upload and manage files from the portal?
Yes. The portal is integrated with the same backend APIs used by the platform for file handling, so users can upload, browse, and manage files through the UI.
If I want files to be vectorized for retrieval or RAG, is a normal file upload enough?
Uploading through Portal Files or POST /v1/files/upload-stream starts background indexing. A successful upload means the file was accepted, not that retrieval is ready: wait for Ready for search in Files, which is upload_stage available without upload_failed in the API. A yellow Ready for search badge means the file is ready but some of its content, such as figures, tables, or picture descriptions, was skipped. Placing an object directly in S3-compatible storage does not start indexing.
For developer integrations, do not call Docling directly from your application code. Use the BSQAI API as the platform integration boundary.
What is Milvus used for?
Milvus stores embeddings and supports similarity search. It is used when the platform needs semantic retrieval, vector search, or retrieval-augmented generation over indexed content.
Workflows and automation
What is the difference between Argo Workflows and Temporal?
They serve different purposes:
Argo Workflowsis used for infrastructure-oriented, Kubernetes-native, and batch-style workflowsTemporalis the preferred workflow engine for developers building application-level and durable business workflows
In practice, platform users and operators may interact with Argo-based workflows, while developers are generally encouraged to build new product or application workflows with Temporal.
Can users create workflows in the platform?
Users and platform teams can work with infrastructure or processing workflows exposed through the platform, especially where Argo Workflows templates or operational flows already exist.
For new application logic, service workflows, or long-running business processes, developers should prefer Temporal.
Can I build workflows without writing code?
Source-to-destination transfers can be configured and scheduled under Data Ingestion in Studio, using connectors registered for your tenant. But fully custom workflows usually still require engineering work, especially when they involve application behavior, integrations, or custom business logic.
How do I monitor workflow progress?
That depends on the workflow type:
Argo Workflowsprovides execution status for Argo-based jobsTemporalprovides visibility for Temporal-managed workflows- platform observability tools such as logs, metrics, and traces help with deeper troubleshooting
AI capabilities
What is KServe used for?
KServe is the model-serving control plane used in BullSequana AI. The BSQAI API creates LLMInferenceService workloads from approved runtime profiles and keeps their endpoints private behind the BSQAI API and the internal inference gateway.
What are LLMs used for in BullSequana AI?
LLMs are used for use cases such as conversational AI, summarization, question answering, extraction, agentic behavior, and other application features exposed through the portal or APIs.
What is RAG?
RAG, or Retrieval-Augmented Generation, combines model generation with retrieved content from indexed data sources. It is useful when answers should be grounded in documents, collections, or enterprise knowledge instead of relying only on the model’s pretrained knowledge.
Can I use my own documents for AI search or question answering?
Yes, but the documents need to go through the correct ingestion and indexing flow so they can be processed, embedded, and made available for retrieval. The portal can participate in that process through its backend integrations, and programmatic integrations can do the same through the BSQAI APIs.
Usage and limits
How do I see how much I use?
Open My usage in the Portal's account menu. It shows your model and API usage in the active organization, in the Portal and through your API keys, and your monthly BCU cap. See Review your usage.
Why is my request refused with "Your monthly BCU cap is reached"?
A tenant administrator can set a monthly cap in BCU on each member, and on the group that funds an API key. Once a cap is reached, new model requests it covers are refused with 429 usage_cap_reached until the month ends or an administrator raises or removes the cap. Retrying the request does not help. See Manage monthly BCU allowances and caps.
Long conversations
Does the model always receive the entire conversation?
No. When conversation history approaches the selected model's context limit, the platform compacts older turns so the conversation can continue. With a configured summarizer model, older context can be summarized; without one, the oldest turns can be dropped from the model's input. The instructions of an active skill are kept in full. The visible transcript is not a guarantee that every detail remains in the next request.
Restate important constraints or attach the relevant source again when an older detail matters. Start a new conversation for a separate task instead of relying on a very long thread to retain exact wording.
Security and permissions
Is data secured on the platform?
The platform is designed with production security in mind, including controlled access, protected service exposure, secret handling, and platform-level security components. The exact security posture still depends on the deployment configuration and the customer environment.
Can access be restricted by user or team?
Yes. Access is controlled through the platform identity and authorization model, including roles, groups, and service-level permissions where applicable.
Support
Who should I contact if something is not working?
Start with your platform administrator or the support contact defined for your deployment. For operational issues, the relevant portal, workflow, API, or platform logs may also need to be checked by the platform team.