Loading
Loading
An internal assistant that answers staff questions from your own SOPs, policies, and project history. Every answer cites its source, permissions mirror your existing document access, and anything uncertain or sensitive routes to a named human owner.
Solution blueprint: a system Callimachus Forge can architect for your operation. Not a case study, and not a claim of past deployment.
At a glance
Stack candidates
Slack · n8n · Postgres with pgvector · Google Drive · Notion · Confluence · Airtable
System run
A simulated run of the exact steps documented on this page, from first trigger to final handoff. Nothing here is live client activity.
Simulation · Loops
The problem
Every operations team runs on knowledge that lives in two places: documents nobody can find and heads nobody can always reach. Staff interrupt senior people for answers that are technically written down somewhere. Different people give different answers to the same policy question. Outdated versions of procedures circulate because they are easier to find than current ones. When an experienced employee leaves, the answers leave with them. Generic chatbots do not fix this: they invent plausible answers, they cannot cite your actual documents, and they have no concept of who is allowed to see what. The result is a slow tax on everyone: repeated questions, inconsistent answers, quiet mistakes.
This blueprint fits organizations whose working knowledge is spread across Google Drive folders, Notion or Confluence spaces, HR policy PDFs, and years of project archives. Slack is where questions actually get asked. SOPs exist, but they were written at different times by different people and stored wherever made sense that week. Access matters: salary bands, legal matters, and security procedures cannot be visible to everyone, and that concern is often the stated reason knowledge stays locked in individual heads. The assistant sits inside this reality rather than replacing it. Existing storage stays where it is, and existing permissions stay authoritative.
The manual process
The automation opportunity
The opportunity is an assistant inside your chat tool that answers questions the way your best operations manager would: from the actual approved documents, with a link to the exact source, and an honest handoff to the right person when it is not sure. Behind the scenes, the system keeps a permission-aware index of your SOPs, policies, and project history, refreshed automatically as documents change. It only shows people material they already have the right to see. When it cannot answer, the question goes to a named owner instead of dying in a channel, and the pattern of unanswered questions becomes a concrete list of documents worth writing. Nothing is invented. Everything is traceable.
Architecture
Every step is typed: triggers, AI decision points, business logic, actions, data, and the human checkpoints that keep judgment where it belongs.
AI components
Human checkpoints
An employee asks the assistant a question via Slack DM or slash command. An n8n webhook receives the message along with the sender's verified Slack user ID and channel context.
n8n maps the Slack user to their directory account and pulls role and group membership from the staff roster. The result is an access filter listing which document collections this person may see.
The question is embedded and run against a vector index of chunked SOPs, policies, and project documents. Metadata filters restrict candidates to collections in the user's access filter before ranking.
The model drafts an answer using only the retrieved passages, citing each source inline. If the passages do not cover the question, it says so explicitly instead of filling gaps from general knowledge.
Rules check retrieval scores, the model's own coverage flag, and a sensitive-topic list (HR, legal, security). Strong answers pass through. Weak or sensitive ones route to the escalation queue instead of the employee.
Escalated items post to the owning team's Slack channel with the draft and sources attached. The owner approves, edits, or replaces the answer, and their response is sent to the asker and captured in the log.
The approved or high-confidence answer returns in the original Slack thread with links to the exact source documents and a thumbs up or down prompt for the asker.
Question, answer, sources, confidence score, escalation path, and feedback are written to a Postgres log table. This is the audit trail and the raw material for gap analysis.
A scheduled n8n job re-crawls Google Drive, Notion, or Confluence, re-chunks changed documents, refreshes embeddings, and updates permission metadata so the index mirrors current source access.
A scheduled model pass clusters escalated and unanswered questions from the log into themes, producing a short list of documents that do not exist yet but keep being asked for.
The themes land as a weekly report to the operations lead: what people asked, what could not be answered, and which SOPs to write next. Documentation work gets prioritized from real demand.
Engineering notes
Benefits are qualitative by design. We do not publish numbers we cannot verify.
The free audit maps your actual tools and process against this architecture, and tells you honestly what is worth building.