# Partial streams

connectPartialStream is for sources where a single row grows over time, like token-streamed assistant replies.


Use `connectPartialStream` when a single row grows over time. The canonical case: an assistant chat row whose `content` field accumulates token by token; `tokens` and `latencyMs` updated each chunk.

## Signature

```ts
function connectPartialStream<
  TRow extends Record<string, unknown> & { id: string },
>(
  grid: GridLike<TRow>,
  stream: AsyncIterable<Partial<TRow>>,
  options: PartialStreamOptions,
): StreamConnection;

interface PartialStreamOptions {
  rowId: string;
}
```

Each yielded `Partial<TRow>` is merged into the row keyed by `options.rowId` via `applyTransaction.update`. The row must already exist (call `applyTransaction.add` once before connecting, or seed it through your engine state).

## Recipe — chat completion

```tsx
import { connectPartialStream } from "@pretable/stream-adapter";

// Seed the row first.
grid.applyTransaction({
  add: [
    { id: "msg-001", role: "assistant", content: "", tokens: 0, latencyMs: 0 },
  ],
});

connectPartialStream<ChatRow>(
  grid,
  openai.responses.stream({ model: "gpt-5", input: prompt }),
  { rowId: "msg-001" },
);
```

Each chunk's `Partial<ChatRow>` (whatever fields it carries — typically `content` deltas accumulated client-side) is merged into the seeded row. The grid's selection, scroll position, and focus state stay anchored to that row across every patch.

## Lifecycle

Same as `connectElementStream` — `done` resolves when the iterator finishes, `dispose()` stops consuming early.

## When you have raw JSON

If your source emits raw JSON strings (e.g. a single object whose keys arrive over time), pipe through `parsePartialStream` first. See [Parsers →](/docs/streaming/parsers).