Knowledge Retrieval
28% faster CS resolution
DTC Shopify Subscription Brand · DTC Shopify Subscription
Client Context
Before the system existed
The customer service team had grown past 30 people. Nearly 1,000 macros, response guides, and policies governed how they replied to customers, and the rules were constantly changing.
There was no central database. Information lived across standard operating procedures, announcements in channels, and whatever happened to be pinned where. When a new question came up, the rep asked. As the team kept expanding, the same questions came up again and again.
“What's the current policy on this, again?”
The Challenge
Why this couldn't be solved by more people or an off-the-shelf tool
Repeated questions are a tax that compounds as a team grows. Each one interrupts whoever happens to know the answer. Each one trains the team to ask first instead of search first. And each one comes back next month, because the underlying policies keep changing.
Without a single source of truth, the team had no way to know whether the answer they were getting reflected this week's policy or last quarter's. The policies were standardized in the heads of the longest-tenured agents, not in any system the rest of the team could query.
Build vs. buy
Why a central database, not another SOP doc
Documentation alone doesn't solve the problem if it lives in five places and nothing makes them consistent. Centralizing the source of truth and exposing it through one queryable interface keeps the answers current and the team aligned, without forcing anyone to learn a new tool.
The Approach
The system, in plain English
A central knowledge base, queryable in plain English. Every macro, response guide, and policy is standardized and indexed. The team asks a question the way they'd ask a coworker; the system returns a plain-text answer grounded in the current source of truth, with a link back to the underlying doc.
Under the hood: a vector index over the standardized corpus, a retrieval layer that ranks passages by semantic relevance, and an LLM that drafts the plain-text answer with citations. If the system can't ground an answer, it says so rather than guessing.
System architecture
Two design choices did most of the work. First, the standardization step, all macros, guides, and policies normalized into a single corpus so retrieval has a coherent surface to search. Second, citations are mandatory, so every answer ships with a source link the agent can verify in one click.
Integrations
How it works for the team
A rep asks: “what's the current policy on international refunds for damaged shipments?”
→ Plain-text answer quoting the standardized policy verbatim, with a link to the source. The rep verifies, replies to the customer, and never leaves their workflow.
A new hire asks: “which response guide applies to a subscription canceled mid-cycle?”
→ The relevant response guide is returned in plain text, with the exact approved wording. No senior agent gets pinged.
An agent asks: “did the carrier-delay macro change this week?”
→ The current macro is returned alongside a note flagging the latest change, so the agent sends the right version, not last month's.
The Results
What changed after launch
Average query resolution time dropped 28%. Questions that used to bounce around the team got grounded, plain-text answers from a single source of truth.
Onboarding got faster. New hires no longer had to absorb tribal knowledge through whoever happened to be online. The current policy lived in one place, queryable in the tool the team already worked in.
| Metric | Before | After |
|---|---|---|
| Source of truth | Scattered across SOPs and channel announcements | One central, queryable knowledge base |
| Answer format | Find the doc, parse the policy, draft the reply | Plain-text answer, sourced, ready to use |
| Query resolution time | Baseline | 28% faster |
| Build timeline | N/A | 5 days |
Outcome
28% faster. Built in five days.
Takeaway
Is this the bottleneck you have?
If your customer service team is past 20 people, your macros and policies live in more than one place, and the same questions keep cycling, you have this bottleneck. A central queryable knowledge base that the team can hit in plain English solves it.
The build itself is fast. The hard part is the standardization, taking what's scattered across docs and channels and turning it into a coherent, indexable source of truth. Once that exists, the retrieval layer on top of it ships in days.
You probably have this bottleneck if…
- 20+ person customer service team
- Macros, response guides, and policies spread across multiple sources
- Policies that change often, with no single place that reflects the latest
- The same questions resurfacing every week
- Onboarding dragging because knowledge is tribal
FAQ
Common questions
How is this different from native chat search or built-in AI?
Built-in chat AI summarizes conversations; chat search returns messages. Neither grounds answers in a curated, standardized source of truth. This system indexes the team's actual macros, response guides, and policies and returns plain-text answers tied to those documents.
Why build instead of using an off-the-shelf knowledge tool?
Off-the-shelf knowledge tools are good when the team can adapt to their structure. Custom is the right answer when you want the index to mirror your own standardized corpus, when you want full control of how answers are drafted, and when you want the system to stay current as the policies change without filing a feature request.
What does the standardization step actually involve?
Taking macros, response guides, and policies as they exist across SOP docs and channels and normalizing them into one coherent corpus. Duplicates collapsed, contradictions surfaced and resolved with the team, structure imposed where the source was loose. The retrieval layer is only as good as the corpus underneath it.
How long does this take to build?
This build was five days from kickoff to live. The variable is how clean the underlying macros, guides, and policies already are. The standardization step is the long pole; the retrieval and answer layer is fast.
Related systems we’ve built
Unified Customer View via MCP Server
How We Replaced a Data Analyst With a Data Analyst: Shopify, Gorgias, Recharge, and Multiple Databases Unified Into One Queryable System
Read case study →Creative Intelligence Pipeline for Meta Ads
How a DTC Growth Team Queries Winning Ad Patterns Across 5,000+ Creatives: Without Looker, Without Spreadsheets
Read case study →Have this bottleneck? Let’s map your version of this system.
One call. A concrete roadmap, whether you build with us or not.