An independent concept project by 4sidedevelop. Screens and data are illustrative.
AI & Knowledge Management
SupportLens
AI assistant for support teams
A support workspace concept where an assistant drafts answers from documentation and past tickets, cites its sources, and leaves the final reply to a person.
- Category
- AI & Knowledge Management
- Proposed stack
- Next.js, TypeScript, PostgreSQL with pgvector, Retrieval over approved documents, LLM API (provider to be selected)
- Status
- Concept — not a client project
01
The challenge
Support agents look for answers across help-centre articles, internal documents, and old tickets. The answer usually exists somewhere, but finding it takes longer than writing the reply.
The team wants help preparing answers without losing control: every suggestion should show where it came from, and nothing should reach a customer without a person reading it first.
02
Who it’s for
Support agent
Works the queue, reviews drafts, checks sources, and sends the final reply.
Specialist
Receives escalated questions with the conversation, the draft, and the sources already attached.
Knowledge owner
Maintains the documentation and sees which questions had no source.
03
The proposed solution
The concept puts the assistant in a side panel next to the conversation. It proposes a draft, lists the documents it used, and says plainly when it cannot find a source. Drafts and sent messages look different, so it is always clear what the customer has seen.
Unified queue
Email, chat, and form tickets in one list.
Knowledge search
Help centre, internal documents, and past tickets together.
Answer drafts
Suggested replies with numbered citations.
Source list
Each citation opens the passage it came from.
Review and escalation
The agent edits and sends, or hands off to a specialist.
Draft feedback
Agents rate drafts; gaps go to the knowledge owner.
04
Key workflows
1Answer with a cited draft
- Step 1: The agent opens a ticket from the queue.
- Step 2: The assistant proposes a draft with numbered citations.
- Step 3: The agent checks the sources, edits the draft, and sends it.
2When no source is found
- Step 1: The assistant cannot find a document that answers part of the question.
- Step 2: It says so and leaves that part out of the draft.
- Step 3: The agent answers from their own knowledge or escalates.
3Escalate to a specialist
- Step 1: The agent chooses Escalate and selects a team.
- Step 2: They add a short note about what the customer needs.
- Step 3: The specialist receives the conversation, draft, and sources together.
4Rate the draft
- Step 1: After sending, the agent marks the draft as used, edited, or not used.
- Step 2: Questions without a source are collected for the knowledge owner.
05
Screens
Illustrative screens with sample data, designed for this concept.
06
Design decisions
Draft and sent never look alike
A draft has a dashed outline and a “Draft — not sent” label. Sent messages are solid and carry a timestamp and the sender.
Citations inline and in a list
Numbers in the draft match the source list, and each source shows its type and when it was last updated.
An honest empty state
When nothing supports a claim, the assistant says what it could not find instead of filling the gap with a guess.
Edits stay visible
After sending, the reply shows what changed from the draft, which makes feedback specific and review easier.
07
What to validate
The concept is a starting point. These questions would be tested with users and resolved during implementation.
- Whether cited drafts save agents time compared with searching manually.
- How agents check citations in practice, and whether the passage preview is enough.
- Which documents are reliable enough to cite; outdated articles need an owner and a review date.
- Topics the assistant must never draft, such as refunds, legal, or security, and how they are routed.
- Which ticket data may be sent to a model provider, and how long it is kept.
Proposed stack: Next.js, TypeScript, PostgreSQL with pgvector, Retrieval over approved documents, LLM API (provider to be selected). The screens on this page were built with React and Tailwind CSS for this demonstration.
Have a similar project?
Tell us about the workflow you want to change. We reply with a clear next step, not a pitch deck.