Many vendors do this: take the money, hand over a login, go quiet. We do the opposite — agree in writing, clean the material, prove it on a small sample, and make sure your own people can add documents and check where an answer came from. Delivery is not a file. It is capability you keep.
The three failures we design against
- Login, then silence. Nobody is taught how the system looks or works; staff keep asking each other.
- Fluent but quietly wrong. The same question has three answers inside the company, and the AI picks the one that is hardest to defend.
- One person can use it. When that person leaves, the system becomes a folder nobody dares to touch.
Our process is not an internal efficiency sheet. It exists to close those three holes.
Six steps, roughly 7–14 working days
| # | Step | What you get | Acceptance |
|---|---|---|---|
| 1 | Scoping (0.5–1 day) | A 30–60 minute call: who uses it, what they will ask, who owns a wrong answer, how it is handled today. We produce a Requirements Sheet with 20 high-frequency questions and what "correct" means for each. | You confirm three sentences: who, asks what, and what counts as correct. Gate: no confirmed scope, no quote, no start. |
| 2 | Inventory & cleaning (1–3 days; ~2 hours of your time) | A Document Inventory & Gap List: what is usable, what needs one authoritative version, what must stay out of the system. | You confirm which version is authoritative; every gap has an owner and a date. Gate: unsettled versions stop the project. |
| 3 | Feasibility sample (0.5 day) | 10–20 core documents turned into a working sample, tested against the 20 questions, result reported item by item. | ≥70% answered correctly, and every critical question correct. Those 20 questions are frozen as Question Set v1, so every later re-test uses the same ruler. | Gate: pass, or we fix material or narrow scope — we do not push on regardless.
| 4 | Build (3–7 days) | An answerable system (web / WeCom / internal tool), admin account, and a Canonical-Ruling Document: when sources disagree, newer beats older, formal documents beat chat logs. | Full test run against step 1's standard, every case logged. Gate: answers carry sources; when the material has no answer, it says so. |
| 5 | Acceptance & handover training (1–2 days) | Two training sessions (staff: how to ask; admins: how to add material), a screenshot manual + top-20 FAQ, account and backup handover. | Four tiers: ① answers correctly ② every answer traceable ③ you add a new document yourself and it takes effect ④ four metrics no worse than the pre-project baseline — accuracy, confident-wrong rate, correct-refusal rate, over-refusal rate — each with a 95% confidence interval. All four, or it is not delivered. |
| 6 | Aftercare (monthly, in service term) | The question set from acceptance is re-run monthly against the four metrics (the confident-wrong rate especially — the answer that sounds right and is not), new material synced, quarterly review; policy, price and qualification changes updated within 24 hours. | Monthly four-metric result not below the acceptance level (same questions, same method, comparable month to month). Two consecutive months below: free remediation, annual fee refunded pro rata. |
What you physically receive
| # | Deliverable | When |
|---|---|---|
| 1 | Requirements Sheet (with the 20 questions) | End of step 1 |
| 2 | Document Inventory & Gap List | End of step 2 |
| 3 | Sample Test Report (accuracy and causes of failure) | End of step 3 |
| 4 | Working system + admin account | End of step 4 |
| 5 | Canonical-Ruling Document | End of step 4 |
| 6 | Screenshot manual + top-20 FAQ | End of step 5 |
| 7 | Handover confirmation (accounts, backup, data ownership) | End of step 5 |
| 8 | Quantitative Acceptance Report (four metrics before/after, with 95% confidence intervals, plus the question set) | End of step 5 |
| 9 | Monthly accuracy report and change log | During the service term |
How quoting works: document inventory first, then a problem list and plan, then a quote. No price is published, because a knowledge-base project's price depends on volume, structure and integrations.
For pricing, email savant0196@126.com.
Acceptance is not a feeling
After remediation, how much better did it actually get? We show it with four metrics and their confidence intervals, not with adjectives.
| Metric | What it answers | Why you should watch it |
|---|---|---|
| Accuracy | Of 30 real questions, how many were answered correctly | The overall "is it usable" number |
| Confident-wrong rate | Answered wrongly, but in a tone that sounds certain | The dangerous one: a colleague acts on it, and that costs more than "I don't know" |
| Correct-refusal rate | When the material has no answer, it says so | Decides whether it will make commitments on the company's behalf |
| Over-refusal rate | It says "I don't know" while an answer exists | Decides whether staff abandon it after two tries |
The method is simple: the sample step already fixes your real questions into one question set; before handover we run it again; you get both sets of numbers side by side, each with a 95% confidence interval — same questions, same method, comparable every month.
We report only what we measured. No promised uplift percentages, no padding with questions outside your business; the report states how many questions were used, what the material covers, and how answers were judged. What we measured is what you see.
FAQ
What do you need from us before starting?
Where the documents live (contracts, quotations, SOPs, product material, support logs), and one person who can confirm which version is authoritative.
What if the sample fails?
We say so. Either we fix the source material or we narrow the scope, and we discuss it before any full build. We do not build on a base that did not pass.
Can our own people maintain it afterwards?
That is the point of step 5. Acceptance is not complete until you add a document yourself and see it take effect.