Customer Service Q&A

Customer service Q&A: cited answers, honest handoffs

Support knowledge rarely lives in one place. Product specifications sit in the latest manual, return policies in an internal rules document, a specific customer's arrangement in a ticket thread, and price or payment terms with the commercial team. An agent handling one enquiry jumps between four systems, pastes the same paragraph again, and still gives an answer that may not match the last one. New hires have it worse: they know the answer exists, not which words to search.

RatioMax puts that material into a knowledge base you can question. Hybrid retrieval with a ReRank pass finds the relevant passage first, and the reply comes back with a clickable citation that opens the manual page or the policy clause it came from. When the documents do not cover something, the system says so and hands the conversation to a person. It does not commit on your behalf or authorise refunds; it answers from what is on file.

Where it breaks

  1. Knowledge spread across several systems

    Manuals, policy documents, price sheets and ticket history sit on shared drives, in internal tools and in chat threads. Handling one enquiry means switching between four windows to find the material, then working out whether what came back is still in force.

  2. One question, two different answers

    Return conditions, warranty periods and who pays freight attract a different answer from each agent. When a customer puts two screenshots side by side, the issue stops being the product and becomes the consistency of your service.

  3. Answers without a source

    Agents answer from memory. When a customer asks which clause says so, the search starts again, and if a dispute follows there is no way to show which version of the policy was in force at the time.

  4. Onboarding by word of mouth

    A new agent shadows a senior colleague, then learns by taking live enquiries. Which replies are standard and which must be escalated is picked up from whoever sits nearby, so coverage collapses whenever the team is short-handed.

How RatioMax responds

  • One searchable library

    Manuals, policies, scripts and past tickets are filed in one knowledge base with metadata for product line, language and effective date. Feishu, Notion, Yuque, RSS and the API can all feed it, and dual-engine parsing turns what arrives into searchable text.

    Platform capability: Multi-source ingestion

  • Answers you can check

    Hybrid vector and keyword retrieval with a ReRank pass narrows the candidates, the model answers from those passages, and every conclusion carries a clickable citation back to the manual page or policy clause. When the material is silent the system says so instead of producing a plausible answer.

    Platform capability: Traceable citations

  • Inside the channels agents already use

    WeCom, Feishu, DingTalk, Yunzhijia, WeChat, QQBot, Slack, Telegram and Mattermost all connect, alongside a mini program, a Chrome extension and a website widget. Agents do not have to switch to another tool, and the same approved answers travel with the channel.

    Platform capability: Channel coverage

  • Scope control per team

    Spaces follow business lines and four-level RBAC narrows access down to a single knowledge base, so in-house agents, outsourced teams and pre-sales each see their own scope. Every access lands in a per-space audit log, identities come from OIDC single sign-on, and two-tier API keys scope integrations.

    Platform capability: Space-level RBAC

Deployment scenarios

  • Pre-sales specification questions

    A customer asks whether a model fits the equipment they already run. Pre-sales asks directly and gets an answer with the manual page and a comparable past project attached.

  • After-sales policy checks

    Return conditions, warranty scope and freight get answered against the policy in force, and the citation can be forwarded to the customer, which removes most arguments about what was said.

  • Agent assist while typing

    The agent asks in WeCom or the web widget, receives a suggested reply with its source, and sends it after their own check. Wording and any commitment stay with the person.

  • Self-service on external channels

    Routine questions are answered by the channel bot first. What it cannot resolve is handed to a person under your rules, with the material it already retrieved passed along so nobody repeats the description.

Terms and knowledge entities

Intent recognition
Working out what an enquiry is actually asking: a specification, a policy, or a complaint. Get it wrong and retrieval and routing both go wrong with it.
Multi-turn dialogue
Successive follow-ups inside one conversation where earlier context must carry forward, otherwise every turn falls back to a generic answer to the first question.
Knowledge entry
The smallest citable unit of knowledge: one entry answers one question and carries a source and an effective date, so it can be revised or retired on its own.
Human handoff
Passing a conversation to an agent when the system cannot answer, is not confident, or the customer asks. The handoff should carry what was already retrieved, not make the customer repeat it.
Approved response script
Reviewed wording for typical situations, including what may and may not be promised. It is versioned knowledge, not a personal habit of the agent.
Service ticket
The record of one customer issue from intake to closure, with category, owner and resolution. It is the raw material for finding recurring problems and turning them into answers.
Conversation archiving
Retaining conversations for quality checks, dispute evidence and training. Scope and retention period belong in a legal review first, not in an ad-hoc decision.
Channel bot
A question-answering entry point inside WeCom, Feishu or a web widget. It speaks for the company to customers, so its answer scope and escalation rules have to be settled in advance.
Retrieval-augmented generation
Retrieval finds evidence in your own documents first, then the model writes the answer from that evidence. Compared with answers drawn from model parameters alone, the result can be checked and owned.

These definitions are also exposed structurally so search engines and AI engines can read and cite the concepts accurately.

Solution questions

How is this different from a keyword chatbot?
A keyword bot matches fixed question-answer pairs and fails as soon as the phrasing shifts. This system retrieves first and generates second, so questions can be asked freely, answers come from your own documents, and each one carries a clickable source. Adding knowledge means updating a document, not authoring a new rule per phrasing.
What happens when it cannot answer?
It says clearly that the material does not cover the question rather than inventing a plausible reply, then hands the conversation to a person under your rules. The handoff carries the candidate documents already retrieved, so the agent continues instead of restarting, and the unanswered question becomes a signal for what to document next.
Does it connect to WeCom, Feishu and DingTalk?
Yes. WeCom, Feishu, DingTalk, Yunzhijia, WeChat, QQBot, Slack, Telegram and Mattermost are all supported, plus a mini program, a Chrome extension and a website widget. Channels are entry points only: one knowledge base and one permission model sit behind them, so every channel returns the same approved answer.
Our content lives in Feishu and Yuque. Can it sync?
Yes. Feishu, Notion, Yuque, RSS, the Chrome extension, the API and the CLI all work as sources, and ingestion parses what arrives. More than ten formats are covered, including PDF, Word, Excel and images: complex-layout PDFs go through OpenDataLoader, scans and images through PaddleOCR-VL.
Can customer data sit on a public cloud?
For most companies the answer is no, and it does not have to be. RatioMax supports full on-premise deployment on Docker Compose or Kubernetes with four core services that scale independently, so data stays inside your network. Each knowledge base can bind a private OSS or S3 bucket, credentials are encrypted with AES-256-GCM, and outbound requests pass SSRF protection.
How do you scope what agents and outsourced teams can see?
Spaces separate teams and four-level RBAC narrows access down to a single knowledge base, so in-house agents, outsourced teams and pre-sales each see their own scope, with every access written to a per-space audit log. OIDC centralises identities and two-tier API keys scope integrations, so removing someone is one action.
Will it invent a policy clause?
Answers are generated from retrieved passages only, and each conclusion carries a citation. When retrieval finds nothing, the system reports that instead of filling the gap. Langfuse and W3C cross-service tracing keep the retrieved passages and prompt for every answer, so a later review can see which policy version was in force.
How do we prove what an answer was based on?
Every answer records its citations, the passages it matched and the model used, and the trace can be replayed end to end. Where conversations are archived for compliance, the reply and its evidence export together, which settles a customer dispute far better than a verbal explanation.

Other solutions