# Stackline AI Full AI Assistant Context

Stackline AI is a provider-neutral TypeScript foundation for AI applications.

Repository:

https://github.com/alexandroit/ai

Public demo:

The repository includes `examples/local-demo`, a safe Vite demo that uses a
local in-memory provider and local in-memory RAG documents. It has no Ollama API
key, no PostgreSQL, no SQLite memory, no private database, and no secret
environment variables.

Package map:

- `@stackline/ai`: provider-neutral contracts, `createStacklineAIServer`, RAG
  flow, memory capture, direct RAG answers, fallback RAG answers, model listing.
- `@stackline/ai-server`: `createStacklineAIHttpHandler` with health,
  manifest, model listing, chat, CORS, request size limit, allowed model policy.
- `@stackline/ai-ui`: `<stackline-ai-studio>` web component with model picker,
  language picker, safe Markdown/HTML rendering, localStorage history, clear
  button, RAG evidence display, and customizable CSS parts/variables.
- `@stackline/ai-ollama`: Ollama provider adapter for local Ollama, Ollama
  Cloud, or an Ollama-compatible backend. Keep API keys on the backend.
- `@stackline/ai-memory-sqlite`: SQLite/sql.js memory store for development and
  small private demos.
- `@stackline/ai-rag-postgres`: read-only PostgreSQL RAG retriever using
  parameterized SQL and row mapping.

Frontend minimal usage:

```ts
import "@stackline/ai-ui";
```

```html
<stackline-ai-studio></stackline-ai-studio>
```

Backend minimal usage:

```ts
import { createStacklineAIServer } from "@stackline/ai/server";
import { createStacklineAIHttpHandler } from "@stackline/ai-server";
import { ollamaProvider } from "@stackline/ai-ollama";

const ai = createStacklineAIServer({
  provider: ollamaProvider({
    target: process.env.OLLAMA_TARGET || "http://127.0.0.1:11434",
    apiKey: process.env.OLLAMA_API_KEY,
    model: process.env.OLLAMA_MODEL || "auto",
  }),
  rag: false,
  memory: false,
});

export const handleAI = createStacklineAIHttpHandler({
  server: ai,
  basePath: "/api/ai",
});
```

RAG design:

RAG belongs to `@stackline/ai` server configuration, not to provider adapters.
Every provider receives a normalized chat request. If RAG is enabled, the
server retrieves contexts and prepends a provider-neutral system message.

Memory design:

Memory stores implement `StacklineMemoryStore`. RAG evidence is display-time
metadata and is not persisted by default. Explicit audit opt-in is required to
store retrieved contexts or RAG evidence metadata.

Security:

Never expose provider credentials, database credentials, RAG SQL, memory paths,
or tenant filters in browser code. The UI should call only the application
backend route.