Package net.thevpc.naru.api.model

Interface NaruStreamChunk

All Known Implementing Classes:
NaruStreamChunkData
public interface NaruStreamChunk
One normalised piece of model output.

This is the only shape the actor loop, the session writer and the renderers are allowed to see. Providers hand over wildly different payloads -- a dedicated reasoning field, an SSE frame, a JSON line, a whole finished body -- and all of it is normalised to this before anything downstream runs, so a renderer never has to know which provider it is displaying.

A chunk is a fragment, not a message. A streamed answer arrives as many ANSWER chunks that concatenate to the answer, and a single response may interleave THINKING and ANSWER freely. Anything that needs the whole thing must accumulate; see index() for the order to accumulate in.

  • Method Summary

    Modifier and Type
    Method
    Description
    boolean
    complete()
    Whether NARU knows this chunk is final, meaning no further content of the same segment can still arrive.
    NaruThinkingExtraction
    extraction()
    How thinking content was recovered, meaningful only for NaruChunkKind.THINKING chunks; null otherwise.
    int
    index()
    Position of this chunk in its response, counting from zero and increasing by one per emitted chunk.
    NaruChunkKind
    kind()
    What this content is.
    String
    provider()
    Provider that produced this chunk, for provenance.
    String
    text()
    The content, which may be empty but is never null.

  • Method Details

    • kind

      NaruChunkKind kind()
      What this content is.

    • text

      String text()
      The content, which may be empty but is never null. For NaruChunkKind.TOOL_CALL this is the call's rendered form rather than a structured payload; structured tool calls remain on NaruResponse.

    • index

      int index()
      Position of this chunk in its response, counting from zero and increasing by one per emitted chunk. Order is what lets a consumer reassemble, so it is assigned at emission rather than derived later.

    • provider

      String provider()
      Provider that produced this chunk, for provenance. May be null for a chunk synthesised by NARU itself rather than by a model.

    • extraction

      NaruThinkingExtraction extraction()
      How thinking content was recovered, meaningful only for NaruChunkKind.THINKING chunks; null otherwise.

    • complete

      boolean complete()
      Whether NARU knows this chunk is final, meaning no further content of the same segment can still arrive.

      This is deliberately conservative, because a streaming reader emits before it knows what comes next. Text that precedes a delimiter this parser has already seen is final; text emitted speculatively mid-stream is not, and neither is text from a reasoning block that never closed.

      The consequence worth knowing: a streamed response reports false on its trailing chunks, and a non-streamed one reports true throughout, because in the batch case the whole body is known before anything is emitted. Callers that need to know whether generation actually finished should ask the source of truth for the whole stream rather than reading this off the last chunk -- see NaruThinkingTagParser.isSegmentComplete().