Buy a RAG server: document search with staff access permissions
The employee asks how to coordinate the change of the project, and receives an answer with the necessary paragraph of the internal regulations. The contractor asks a similar question and sees only the instructions allowed. This is what a useful corporate search looks like: a common interface, but different sources available. To buy a GPU server AI Server Manager should describe this workflow, and then choose the equipment and application for documents, employees and load.
Start with a library that is actually used.
Take the project organization: general instructions are available to all employees, working documents are available to project participants, annexes to the contract are available to a dedicated group. In the first stage, one project and several typical issues are enough. The result should save time on the search: a short explanation, the title of the document, the editorial board, the page and the transition to the authorized original.
RAG means that the application first finds the appropriate snippets and then transmits them to the models to prepare the response. Therefore, the project includes a server, and downloading documents, and search, and communication with company accounts. A scan requires recognition; a complex table requires saving headings and notes. These are independent works that are useful to include in the quotation.
The rights accompany the document until the answer.
The source needs an owner, project, editorial staff, status and permitted groups. This information is transmitted to the associated fragments of the index. User groups define the server part of the application after logging in, not the field that the employee can arbitrarily change in the request. The search is performed within the allowed set; the context of the model is formed from it.
Qdrant provides search terms for payload fields and their logical combinations. This is a tool for implementing the chosen scheme, not a ready-made system of corporate powers. If the document directory is stored in PostgreSQL, its string-level access policies can be used as an additional layer. The appropriate role is chosen for the application: the table owner and privileged roles have a special behavior that the designer considers separately.
| Source | Who needs it? | What the employee gets |
|---|---|---|
| General instruction | Staff of the organization | Response with current version |
| Project work folder | Project participants | Fragments of its project area |
| Annex to the treaty | Appointed group | Permitted item and reference to the original |
| Replaced version | Under the agreed rules of the archive | Visible status to distinguish it from the current |
Access changes – part of daily work
The contractor completed the project, transferred the document to a closed folder or released a new edition. The application needs a consistent order of updating the index, groups and saved results. For the period between the change of rights and the full processing of the archive, a temporary exclusion of the affected source may be provided. So the employee receives an understandable status update, and the administrator – a managed process.
The rules also apply to the history of dialogs, unloading and general cache. For example, a saved manager’s response cannot be automatically shown to a contractor just because the question matched. It is convenient for the user to see the date of the response and the used edition, and after updating the document - to get a new search for the current set.
Select the server by mode of use
For quotation, source volume, scan percentage, daily growth, number of simultaneous employees and desired response time are important. The primary load of the library and the usual working search create different loads. You can agree on night indexing and dedicated resource for daytime questions, while maintaining access to already prepared documents.
CPU, RAM, drives, accelerators and software environment are selected according to the agreed model and control library. The question of purchase - what results the department gets in its working mode. The number 256 in the quotation designation does not replace the memory specification and does not determine the number of users.
It is helpful for a pilot to prepare one question for the employee, manager and contractor, then repeat it after transferring the document to another group. Together, we evaluate the usefulness of the pages found, the accuracy of the editorial board and the actual response time. Such a small, complete process makes it possible to expand the library to other departments.
Server and implementation by one agreed project
Main own quotation - AI Server Manager 256. Price, availability, actual configuration and composition of works are confirmed by quotation. Include an example of documents, access groups, number of employees and city of delivery. Downloaders, interface, deployment, instructions, and maintenance are coordinated with the equipment.
Document enters search
- Original, editorial and owner
- Access groups and status
- Fragments with the same connection to the source
The employee asks a question.
- Entry and relevant groups
- Search for permitted fragments
- Reply with the transition to the original
Changing rights updates search and saved results. Opening the original rechecks the access.
Questions before ordering
Do I need to retrain the model on all the company documents?
The described RAG first searches for sources and prepares a response to the fragments found. Training is not a mandatory part of this scenario. Model composition, search and quality in your library are agreed separately.
Can we start with one department?
Yeah. Convenient first stage – limited library, several groups and clear working questions. The download architecture and rights are designed with the subsequent expansion in mind.
Are rights settings and documents included in the server price?
The composition of the work is fixed in the quotation. It can include integration of accounts, downloaders, search application, group setting and maintenance. These jobs cannot be automatically considered included.
The server in our directory
AI Server Manager 256
Price on request. Actual CPU, RAM, GPU, accelerator memory, drives and availability are determined by quotation. Equipment and implementation are consistent in composition.
Request server and composition of workPrimary sources: Qdrant Filtering; PostgreSQL Row Security Policies. The documentation describes the components of a possible application; their inclusion in the delivery and operation on the selected machine are agreed separately.