Turn work instructions into a source-grounded assistant

Start with the operational bottleneck

Work instructions often describe a process more precisely than the vocabulary employees use to ask about it. Someone asks how to hand over a job, while the relevant document is filed under service completion. An assistant can bridge that language gap if its answers remain tied to approved material.

Start with a small collection of instructions that have clear owners and revision dates. Use real question patterns collected from the team, with sensitive details removed. Avoid beginning with a company-wide ingestion of every folder: an uncontrolled corpus makes both evaluation and maintenance harder.

Design the target workflow

Maintain an ingestion ledger: source identifier, revision, checksum, ingestion date, and access rule. Retrieve passages using both semantic similarity and known identifiers such as process codes. Supply the model with the relevant passages and a concise instruction to cite rather than invent.

An answer should expose the applicable source and state when the procedure does not cover the request. For multi-step instructions, preserve the original ordering and approval points. Record unanswered questions as a backlog for the document owner. Do not let user feedback directly rewrite the source material or train the model without review.

Connect to the existing systems

Connection	Information	Control
Procedure library	Approved text and revision metadata	A document owner publishes changes
Retrieval index	Passages and permission tags	Refresh changed documents; remove withdrawn text
Team interface	Question, answer, citations and feedback	Keep a direct link to the source

Use stable identifiers and versioned inputs. A retry should resume the same piece of work, with the original permissions and evidence attached. Treat an upstream outage as a visible workflow state.

Fit the model to the task

Compare simple search, retrieval with a small model, and a larger model on the same questions. Fine-tuning is not a substitute for keeping procedures current. A retrieval index can be updated when a document changes without retraining the language model. Choose embeddings and chunk sizes using retrieval tests before changing the generator.

Keep a clear human boundary

Test plausible but unsupported questions, ambiguous process names, and instructions containing distracting text. Require the assistant to abstain when it cannot find the answer. A person remains responsible for acting on an instruction and for identifying when the real situation falls outside it.

Measure the complete operating cost

Measure the complete lookup: question, retrieval, reading, verification, and escalation. Include source maintenance and access administration. Avoid presenting a faster first answer as a completed-task saving when employees still have to search manually to verify it.

Pilot checklist

• Assign owners to a bounded procedure collection.
• Build questions with known answers and source passages.
• Test missing, conflicting, and restricted material.
• Verify source revision and answer grounding separately.
• Review unanswered questions as a documentation improvement queue.

Agree on acceptance thresholds before testing, then compare the proposed workflow with the same work handled through the current process. Expand only after the operation can explain both successful and failed cases.

Further implementation detail

• Choose the lowest-cost model that meets the task requirements
• Serve a local model to a team
• Ollama’s local API reference