Private AI / buying guide
Private AI search for business documents
A useful document search system must find relevant material, respect staff access, and make its answers checkable. Before choosing a model, define which records it may use and how staff will verify the result.
Start with a defined collection
Choose an initial collection with a clear owner, such as approved policies, project records, or standard templates. Record the source, current version, and permitted users. Keep superseded documents distinguishable from current instructions.
Test scanned PDFs and image-heavy files as well as text documents. A page that looks readable to a person may need text extraction or optical character recognition before search can use it. Tables, handwriting, and diagrams should be part of the evaluation when they matter to your work.
Search and answer generation
One common approach retrieves relevant passages and supplies them to a language model alongside the question. This is called retrieval-augmented generation, or RAG. It allows a system to use a document collection without placing every file into each request.
Retrieval and generation can fail separately. The search may select the wrong passage, or the model may misread a relevant one. A source reference helps a reviewer check an answer, but it does not prove that the answer is correct.
Ask the installer to demonstrate how staff open the supporting material and what happens when the collection does not contain an answer. The exact search, citation, and indexing behavior belongs in the agreed configuration.
Access to files and connections
Define access before indexing. Your organization should decide who can use the system and which files each workflow may read. Test restricted collections with the actual staff roles that will use them.
For SPLAY’s standard local configuration, staff access the Node through a browser on their network. Files and model traffic stay local, and connections to external AI models are blocked. Local processing still requires an approved access policy, backups, and ongoing administration.
An existing application is a separate requirement. Word files, Excel exports, and scanned records can be discussed as inputs. Direct access to a document management platform, email system, or other account requires an explicitly scoped connection.
A small evaluation set
Prepare questions with known answers and record the document and passage supporting each one. Include questions that require several records, an obsolete version, and an answer the collection does not contain.
Review whether the system found the right material, used the current version, and kept the answer within what the sources support. Repeat important checks after changing the model or document collection.
- Find a fact in one approved document
- Compare terms across two versions
- Summarize a record while retaining the source references
- Ask a question with no supported answer
- Try a request for material outside the user’s permitted collection
Define the installation
SPLAY selects hardware and a model around your workload. Document collection size, file preparation, concurrent use, and software connections affect the scope. Machines are pre-order only, with configurations and services by consultation only.