Ask approved company knowledge
A private document assistant that shows its sources and its limits.
The target is not a chatbot that sounds confident. It is a controlled retrieval service that finds approved evidence, cites it and declines when support is weak.
Start with one bounded document set and a representative question set. Expand only after retrieval, permissions and answer support pass.
Workload before hardware
Describe the queue, the quality bar and the operating owner.
- 01 / Work unit
- Size from model, context, concurrent users and peak demand.
- 02 / Acceptance
- Agree quality, latency, citation or output checks before buying.
- 03 / Operations
- Plan updates, monitoring, support and a fallback route.
Scope the knowledge base
Choose one business process, source owner, document set and update frequency. Remove duplicates and obsolete material before ingestion.
A smaller controlled corpus is a better pilot than “all company knowledge”.
Demand shape
Peak demand can matter more than the daily average
- Work unit
- Tokens, frames, files or jobs.
- Duration
- How long one active job occupies capacity.
- Concurrency
- How many jobs overlap.
- Deadline
- Interactive response or queued completion.
Trace the work before sizing the capacity.
The queue, memory shape, data path and management route turn a broad workload name into a testable service.
Preserve permissions
Retrieval must not give a user access to a passage they could not open in the source system. Separate collections or group-aware filtering may be required.
Permission design is part of the workload, not a later UI setting.
Acceptance bench
Quality and service conditions pass together
Evaluate with real questions
Measure whether the right passage is retrieved, whether the answer is supported, whether the citation is usable and whether refusal works.
Human review remains mandatory for material legal, financial, clinical or safety decisions.
Operating loop
The workload continues after the first demonstration
- Observe Demand, errors and resource state.
- Review Quality drift, access and incidents.
- Change Versioned model or runtime update.
- Retest Focused acceptance before wider use.
Choose hardware after the test
Daily demand depends on model, context, response length and concurrent questions. Ingestion and embedding can be scheduled separately.
The fit test determines whether Team 32, a larger workstation, a rack server or a hosted route is proportionate.
Questions answered
Straight answers to common questions
Can the assistant read all our company files?
It can ingest approved sources after format, permission, retention and quality review. “All files” is not a safe default.
Will every answer cite a source?
The system can be designed to cite retrieved passages and refuse unsupported questions. The behaviour still needs evaluation.
Is SharePoint integration included?
Not in the baseline package. Source connectors and permission mapping require a separate defined scope.
Continue the decision
Useful next steps
Put the claim to work
Turn this guidance into a testable requirement.
The brief asks about workload and operating conditions - not just budget.