Knowledge assistants and RAG
Source-linked AI assistants for industrial knowledge
At shift handover, an operator asks about a recurring fault. Someone searches a manual, someone messages the experienced colleague, and the team still needs to check which procedure applies. NeoBram builds knowledge assistants that bring approved records and source references into that search, helping people reach a checked answer without repeating the same hunt across folders.
Start small, review the result
A practical first step
- Start with
- One approved document collection, one user group and the questions they ask repeatedly.
- The scoped work can produce
- A scoped trial can produce a collection and access map, a reviewable assistant and evidence-quality findings.
- Your team brings
- Representative documents, recurring questions and answers your experts accept. Include a document owner who can resolve conflicting sources.
- Next decision
- Extend the tested collection, narrow the trial or repair an evidence gap before continuing.
Deliverables, access and responsibilities are agreed for the engagement.
When the same questions keep interrupting the shift
Start with the questions that interrupt a team’s day. A maintenance team may need an equipment manual; a project engineer may need the latest approved specification. A pharma investigator may need prior deviations and controlled quality records. Define the collection and context so similar language from another asset, product, project or revision does not become a misleading answer.
Record where people search, who they interrupt and what happens when two documents disagree. When an answer depends on missing site context, the assistant should ask for it. When the question needs judgment beyond the source, it should hand the evidence to a qualified reviewer.
From a question to evidence someone can use
Retrieval-augmented generation, or RAG, finds relevant passages and uses them to prepare a response. The output might be a search result, evidence summary or draft explanation. The illustrative record below shows what a reviewer should be able to inspect when the sources conflict.
The agreed scope determines file types, languages and access groups. A citation needs to lead to the relevant passage and applicable revision; a document name alone does not establish that an answer is supported.
An answer that leaves a conflict visible
Question: which document applies to the inspection record for example unit A?
Source passage A
“Use inspection record A.”
Example manual · revision 3 · section 4.2, page 6 · marked approved in the supplied set. Locator and passage are synthetic.
Source passage B
“Use inspection record B.”
Example local instruction · revision 2 · section 2.1, page 3 · also marked approved. Its applicability has not been resolved. Locator and passage are synthetic.
Draft answer
Unresolved: the supplied sources disagree.
Both passages stay visible. The assistant does not choose a procedure from conflicting evidence.
Reviewer action
Document owner confirms which source applies.
Keep the question open until the source conflict is resolved; then review the answer again.
This synthetic record shows a review format, not an operating instruction or evidence of a customer deployment.
Related work: plant knowledge, investigations and parts research
NeoBram’s published factory knowledge assistant account describes a source-linked conversational interface across procedures, manuals and issue records. It reports read-only ingestion, role-based access and deployment inside the customer network. This is a relevant example of organizing plant knowledge around operator questions.
NeoBram’s published pharma investigation account describes private retrieval of approved quality records and comparable events. Investigators retained conclusions and approval; the assistant did not approve or close deviations or corrective and preventive actions, commonly called CAPAs. The published offline spare-parts account describes local search across equipment records and vendor documents, with engineering verification of applicability and compatibility.
These are NeoBram-reported anonymized accounts, not independent validation; each linked brief explains the limits of its outcome evidence.
Keep the collection useful as the work changes
Your document owners decide what is current, approved and available to each group. NeoBram designs the search and response workflow around those decisions. During preparation, we identify missing labels, conflicting versions and gaps in the collection. Cleaning every historical folder is rarely the right first project; a smaller maintained collection can be a better starting point.
For a pharmaceutical SOP assistant, the approved collection must identify effective versions, site and product scope, and each user’s permitted access. Answers should show the source and revision; superseded, conflicting or missing procedures should trigger a warning or escalation. Quality owners approve the intended use and tests. The assistant supports access to controlled information and does not approve a procedure or replace required training.
The assistant should appear where the team can use it during the task, with an obvious route to the original document and to report a wrong answer. A private or offline design may fit records that cannot leave the site. Local operation still needs suitable hardware, user access and a controlled way to refresh the collection. A document owner and technical operator need distinct, agreed responsibilities.
Test the evidence before widening the collection
Use questions your specialists have already reviewed, including ambiguous, outdated, conflicting and unanswered cases. Separate a search failure from a missing-document problem. A correct citation, a faithful summary and a useful answer are different checks.
Include representative tables, scans, drawings and cross-references when defining supported inputs. Before handover, test a document replacement and withdrawal, an access restriction and an unavailable import. Agree who receives error reports and how users find the approved fallback.
This RAG solution focuses on retrieving approved sources for a bounded task. Company Expert AI has a broader training-led product scope: an industry-trained small-model foundation, company knowledge and refinement with subject-matter experts. A RAG engagement does not by itself include that full product.
- Retrieval: did the correct approved passage and applicable revision reach the reviewer?
- Answer: is the response faithful, correct and useful for the question, including when it should remain unanswered?
- Access: test permitted and restricted sources separately from answer quality; an accurate response can still expose the wrong information.
- Handover: retain evaluated examples, known limits, source ownership and a route to report a problem.
Give operators and experts time back for the job
Measure the full path from asking a question to reaching a checked answer. Include searching, opening sources, correcting the response and asking a specialist for help. Compare similar questions before and during the trial, and keep unresolved questions in the record. The aim is usable time returned to operators and experts, with evidence quality maintained. Agree whose work should become easier, then use that result to decide whether to expand the collection.
More clarity
Questions and answers
Can it work with our current documents?
We assess representative files, readability, revision information and permissions before confirming the scope.
Does a source link guarantee a correct answer?
No. The system can retrieve the wrong material or summarize it incorrectly. Both steps need testing.
Can this become a company chatbot?
Yes, if chat fits the task. We keep the approved information, user permissions and escalation rules explicit as the scope grows.
Do we need to reorganize the entire archive first?
No. A bounded, maintained collection is often the better first test. Choose the users, questions and approved sources, then resolve gaps that affect those tasks. Wider migration can follow evidence of usefulness.
What should we receive from a first trial?
Agree the collection, access and revision rules, a working evaluation interface, test questions and a report showing incorrect and unanswered responses. Production interfaces, migration and ongoing operation should be explicit in the scope.
Start with one business problem