ollama: run /api/generate as /api/chat when a format is set (0.2.5)

gpt-oss:120b returns an EMPTY response on /api/generate whenever a JSON
format is requested. Verified against forge directly, bypassing this
router, so it is upstream behaviour and not ours:

  /api/generate + format:'json'  -> response ''   done_reason stop
  /api/generate, no format       -> response 'OK'
  /api/chat     + format:'json'  -> content '{"tags":["x"]}'

Karakeep tags via /api/generate WITH format, so every job returned
empty and it reported 'the model ignored our prompt'.

Fix: when a format is set, run the request upstream as /api/chat and
hand the answer back in generate's shape (response/context). Callers
change nothing.
This commit is contained in:
Dolon (Agent)
2026-09-19 13:47:21 +00:00
parent a5bd550971
commit a8b6a3ee6e
+41 -1
View File
@@ -316,6 +316,7 @@ async def _forge_raw(
mode: str,
stream: bool,
strip_thinking: bool = False,
as_generate: bool = False,
) -> Response:
"""forge is Ollama: forward as-is (model pinned), return its bytes untouched."""
client: httpx.AsyncClient = request.app.state.client
@@ -334,6 +335,14 @@ async def _forge_raw(
if path == "/api/chat" and _starved(data.get("message") or {}, data.get("done_reason")):
headers[STARVED_HEADER] = "content-starved-by-reasoning"
log.warning("forge answer had thinking but no content (done_reason=%s)", data.get("done_reason"))
if as_generate and isinstance(data.get("message"), dict):
# Present the chat answer in generate's shape for the caller.
msg = data.pop("message")
data["response"] = msg.get("content", "")
if msg.get("thinking") and not strip_thinking:
data["thinking"] = msg["thinking"]
data.setdefault("context", [])
content = _dumps(data).encode()
if strip_thinking and isinstance(data.get("message"), dict) and "thinking" in data["message"]:
# Caller asked for no thinking; give it none -- without having asked the
# model to stop reasoning, which would have produced no answer at all.
@@ -433,6 +442,30 @@ async def _route(request: Request, mode: str, path: str, body: dict[str, Any], s
# think=false, so this fixes a latent failure there too.
strip_thinking = payload.pop("think", None) is False
# 2026-09-19: gpt-oss:120b returns an EMPTY `response` on /api/generate whenever a
# JSON `format` is requested -- verified against forge DIRECTLY, bypassing this
# router, so it is an upstream model/runtime behaviour and not ours:
# /api/generate + format:"json" -> response "" done_reason "stop"
# /api/generate, no format -> response "OK"
# /api/chat + format:"json" -> content '{"tags":["x"]}' (works)
# Karakeep tags via /api/generate WITH format, so every job came back empty.
# Fix: when a format is set, run it upstream as /api/chat (which works) and hand
# the answer back in generate's shape. Callers need change nothing.
rewrote_generate = False
if path == "/api/generate" and payload.get("format"):
msgs: list[dict[str, Any]] = []
if payload.get("system"):
msgs.append({"role": "system", "content": payload["system"]})
u: dict[str, Any] = {"role": "user", "content": payload.get("prompt") or ""}
if payload.get("images"):
u["images"] = payload["images"]
msgs.append(u)
payload = {k: v for k, v in payload.items() if k not in ("prompt", "system", "images", "raw", "context")}
payload["messages"] = msgs
path = "/api/chat"
rewrote_generate = True
log.info("rewriting /api/generate -> /api/chat (format set; gpt-oss returns empty on generate+format)")
if CONFIG.keep_alive != 0:
# Residency: a cold load is 21-66 s and lands entirely in front of the
# caller. An explicit finite keep_alive from a caller BEATS forge's
@@ -444,7 +477,14 @@ async def _route(request: Request, mode: str, path: str, body: dict[str, Any], s
payload["keep_alive"] = CONFIG.keep_alive
try:
return await _forge_raw(
request, health.upstream, path, payload, mode, stream, strip_thinking=strip_thinking
request,
health.upstream,
path,
payload,
mode,
stream,
strip_thinking=strip_thinking,
as_generate=rewrote_generate,
)
except UpstreamError as exc:
locator.invalidate()