Package net.thevpc.naru.ext.models.ollama

Class NaruOllamaNativeStreamParser

java.lang.Object

net.thevpc.naru.ext.models.ollama.NaruOllamaNativeStreamParser

All Implemented Interfaces:

NaruStreamResponseParser

public class NaruOllamaNativeStreamParser
extends Object
implements NaruStreamResponseParser

Reads Ollama's /api/chat stream.

Not SSE, and the difference matters: Ollama writes one bare JSON document per line with no data: prefix, no frame separator and no [DONE] sentinel. Termination is a done:true field on the last document. Reusing the SSE reader here would silently produce zero events rather than an error, which is the worst possible failure mode -- it looks like a model that has nothing to say.

Reasoning arrives in message.thinking for models that emit it, in its own field and interleaved with the answer, so it is fed as native thinking and kept out of the answer. Content is still scanned for thinking tags, because a model configured for inline reasoning puts the tags there instead -- the same rule the OpenAI-compatible path uses, and the reason a streamed Ollama turn equals a batched one.

Tool calls arrive whole rather than in fragments, so unlike the OpenAI-compatible protocol there is nothing to reassemble here.

Constructor Summary

Constructors

Constructor	Description
NaruOllamaNativeStreamParser(String provider, NaruStreamHandler handler, NaruModelConfig model)

Method Summary

Modifier and Type	Method	Description
NaruResponse	finish(boolean interrupted)
boolean	hasDeliveredContent()	Whether any user-visible output has already been delivered.
boolean	isCancelled()	Polled between documents, which is where a cancel lands.
boolean	onLine(String line)	Consumes one NDJSON document.
NaruResponse	read(net.thevpc.nuts.net.NHttpResponse response)	Consumes the body incrementally and returns the assembled response.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Details

NaruOllamaNativeStreamParser

public NaruOllamaNativeStreamParser(String provider,
 NaruStreamHandler handler,
 NaruModelConfig model)

Method Details

read

public NaruResponse read(net.thevpc.nuts.net.NHttpResponse response)
                  throws IOException

Description copied from interface: NaruStreamResponseParser

Consumes the body incrementally and returns the assembled response.

Must deliver events as they arrive rather than reading the whole body first: the entire point is that a renderer sees text while the model is still generating it.

Specified by:

read in interface NaruStreamResponseParser

Throws:

IOException

onLine

public boolean onLine(String line)

Consumes one NDJSON document.

Returns:

false once the stream has ended

finish

public NaruResponse finish(boolean interrupted)

hasDeliveredContent

public boolean hasDeliveredContent()

Description copied from interface: NaruStreamResponseParser

Whether any user-visible output has already been delivered.

Consulted by the caller before retrying. A stream that has delivered nothing can be retried invisibly; one that has cannot, because the tokens cannot be recalled.

Specified by:

hasDeliveredContent in interface NaruStreamResponseParser

isCancelled

public boolean isCancelled()

Polled between documents, which is where a cancel lands.