AI chatbot or assistant

A model-backed assistant: conversations with streamed responses, retrieval over your own documents, token accounting, and guardrails for when the model is wrong or the input is hostile.

You can change every choice once the builder opens.

Core features for the MVP

Conversation history with streamed responses
Document ingestion, chunking, and embedding into a vector store
Retrieval of relevant passages before answering, with citations shown
Per-user rate limits and a daily budget cap
Token usage recorded for every request
Guardrails: user text treated as data, never as instructions
A visible way for the user to report a bad answer

Left for later

Tool use so the assistant can act, not only answer
Evaluation set with regression scoring on prompt changes
Multiple models selected by task and cost
Human handover for questions the assistant should not answer

Recommended stack

Section	Recommended
Language	TypeScript
Frontend framework	Next.js
UI and styling	Tailwind CSSshadcn/ui
Backend	Next.js server
API style	REST
Database engine	PostgreSQL
Database provider	Neon
ORM and data access	Drizzle
Caching	Upstash Redis
Authentication	Clerk
Authorization and roles	Role-based access control
Security	Input validationRate limitingSecurity headers and CSPSecrets managementAudit logs
File storage	AWS S3
Background jobs and cron	Inngest
AI features	Anthropic APIpgvector embeddings
Hosting and deployment	Vercel
CI/CD	GitHub Actions
Monitoring and analytics	SentryPostHog
Testing	VitestPlaywright
Extras	Accessibility (WCAG)BackupsDocumentation

Data model draft

Conversation

One ongoing exchange between a user and the assistant.

id
userId
title
model
createdAt
lastMessageAt

Belongs to one User
Has many Messages

Message

One turn, from the user or the assistant.

id
conversationId
role
content
citedDocumentIds
inputTokens
outputTokens
createdAt

Belongs to one Conversation

Document

A source the assistant may draw on.

id
title
sourceUrl
storageKey
mimeType
ingestedAt
checksum

Has many DocumentChunks

DocumentChunk

A passage with its embedding, retrieved to ground an answer.

id
documentId
position
text
embedding
tokenCount

Belongs to one Document

UsageRecord

Token spend per request, for cost control and limits.

id
userId
provider
model
inputTokens
outputTokens
createdAt

Belongs to one User