Feedback Agent

AI customer support that turns every conversation into product evidence.

简体中文 · Architecture · API · Deployment

Feedback Agent is a self-hosted, multi-app customer-support and Voice-of-Customer platform. It answers user questions, then turns the same conversation into structured feedback: intent, category, pain points, impact, duplicate signals, draft requirements, owner follow-up, and reports.

It is built for product teams that want one reusable feedback layer across Flutter apps and the web without shipping model keys or app secrets to client devices.

Why teams use it

Product outcome	What the platform does
Resolve common questions	Retrieves app-scoped knowledge and streams cited answers
Capture feedback without another form	Extracts intent, pain points, impact, and resolution state from the conversation
Find repeated demand	Groups similar feedback while preserving original evidence
Keep AI accountable	Creates requirement drafts; a human confirms, rejects, or merges them
Reuse one backend	Isolates configuration, credentials, knowledge, and data by validated app ID
Meet users where they are	Provides a public Next.js entry and an embeddable Flutter SDK

From conversation to action

flowchart LR
  User["User conversation"] --> Intent["Intent and resolution"]
  Intent --> Knowledge["App-scoped knowledge"]
  Intent --> Feedback["Structured feedback"]
  Feedback --> Dedupe["Similarity grouping"]
  Dedupe --> Draft["Human-reviewed requirement draft"]
  Draft --> Queue["Owner queue and reports"]

Loading

LangGraph.js runs inside the NestJS API as a controlled workflow with explicit state and decision boundaries. The project deliberately avoids an open-ended multi-agent conversation for this path: support and feedback processing need predictable latency, traceable evidence, and human approval.

What is included

• NestJS API with authentication, app isolation, Swagger, SSE, and LangGraph.
• BullMQ worker for knowledge ingestion, feedback extraction, deduplication, and report generation.
• PostgreSQL with pgvector, Redis, and S3-compatible object storage.
• Next.js public landing page and bilingual feedback conversation.
• Vue 3 administration console for apps, knowledge, conversations, feedback, requirements, owner queues, model settings, and reports.
• Flutter client, controller, full-page UI, embedded view, launcher, and Android/iOS/Web example.
• Local and production Docker Compose configurations.

Quick start

Requirements: Node.js 22.13–26, pnpm 11, and Docker or OrbStack. FVM and Flutter 3.44.2 are only required for Flutter SDK development.

git clone https://github.com/yjsf216/feedback-agent.git
cd feedback-agent
cp .env.example .env
pnpm install --frozen-lockfile
pnpm infra:up
pnpm db:generate
pnpm db:migrate
pnpm db:seed
pnpm dev

Open:

• Web: http://localhost:3000
• Demo conversation: http://localhost:3000/feedback/demo
• Admin: http://localhost:8848
• API and Swagger: http://localhost:4100/docs
• MinIO Console: http://localhost:9001
• Mailpit: http://localhost:8025

The development seed reads ADMIN_EMAIL and ADMIN_PASSWORD from your local .env file. Empty model keys enable a deterministic fallback for testing the workflow; they do not represent real model quality.

Flutter integration

Use the SDK directly from this repository until a registry release is ready:

dependencies:
  feedback_agent_flutter:
    git:
      url: https://github.com/yjsf216/feedback-agent.git
      path: packages/flutter_sdk

final client = FeedbackAgentClient(
  baseUri: Uri.parse('https://feedback-api.example.com'),
  appSlug: 'my-product',
);

final controller = FeedbackChatController(gateway: client);

Use FeedbackChatPage for a complete route, FeedbackChatLauncher for a floating entry, or FeedbackChatView inside an existing screen. Existing app identity must be exchanged by the host backend; never place an app secret or model key in a Flutter binary. See the Flutter SDK guide.

Production deployment

The production Compose stack binds API, Web, Admin, and the MinIO console to 127.0.0.1; PostgreSQL, Redis, and the MinIO API stay on the Docker network. Put a TLS reverse proxy in front of the public services and never publish the MinIO console, database, or Redis directly.

cp .env.production.example .env.production
# Replace every placeholder and public URL before continuing.
pnpm prod:config
pnpm prod:build
pnpm prod:up
pnpm prod:bootstrap  # first initialization only

Read the deployment and backup guide before using the stack with real feedback.

Security boundaries

• Every domain query is scoped by a server-validated app ID.
• Model keys, signing credentials, and app secrets stay server-side.
• AI-generated requirements remain drafts until a human acts.
• Example credentials are for local development only.
• Do not put production exports, conversations, screenshots with user data, environment files, or private keys in the repository.

Please report vulnerabilities through GitHub private vulnerability reporting, not a public issue. See SECURITY.md.

Project status

Feedback Agent is an early-stage, working MVP. The main support-to-feedback flow, multi-app isolation, Flutter/Web entry points, administration console, and single-host deployment are implemented. Real-time human takeover, scheduled reports, recursive crawling, OCR, voice, message attachments, and external notification webhooks are not implemented yet.

Contributing

Bug reports, integration examples, documentation improvements, and focused pull requests are welcome. Start with CONTRIBUTING.md and the open-source policy. If this project matches your workflow, a GitHub star or a concrete use-case Discussion helps shape the roadmap.

License

Original Feedback Agent code is licensed under the Apache License 2.0. The administration app contains MIT-licensed pure-admin-thin code; its original notice remains in apps/admin/LICENSE. See THIRD_PARTY_NOTICES.md for details.