Skip to content

The open-source AI voice studio. Clone, dictate, create.

Address review: readiness invalidation, redirects, endpoint gating, remote model label

jamiepine129d84f6c703
diff --git a/app/src/components/ServerTab/CapturesPage.tsx b/app/src/components/ServerTab/CapturesPage.tsx
index 415bfc3c..db5bd19e 100644
--- a/app/src/components/ServerTab/CapturesPage.tsx
+++ b/app/src/components/ServerTab/CapturesPage.tsx
@@ -520,7 +520,6 @@ export function CapturesPage() {
value={customLlmEndpointDraft}
onChange={(e) => setCustomLlmEndpointDraft(e.target.value)}
onBlur={commitCustomLlmEndpoint}
- disabled={!autoRefine}
autoComplete="off"
spellCheck={false}
aria-label={t('settings.captures.refinement.customEndpoint.endpointLabel')}
@@ -535,7 +534,7 @@ export function CapturesPage() {
value={customLlmModelDraft}
onChange={(e) => setCustomLlmModelDraft(e.target.value)}
onBlur={commitCustomLlmModel}
- disabled={!autoRefine || !customLlmEndpointDraft}
+ disabled={!customLlmEndpointDraft}
autoComplete="off"
spellCheck={false}
aria-label={t('settings.captures.refinement.customEndpoint.modelLabel')}
@@ -560,7 +559,7 @@ export function CapturesPage() {
value={customLlmApiKeyDraft}
onChange={(e) => setCustomLlmApiKeyDraft(e.target.value)}
onBlur={commitCustomLlmApiKey}
- disabled={!autoRefine || !customLlmEndpointDraft}
+ disabled={!customLlmEndpointDraft}
autoComplete="new-password"
spellCheck={false}
aria-label={t('settings.captures.refinement.customEndpoint.apiKeyLabel')}
diff --git a/app/src/lib/hooks/useSettings.ts b/app/src/lib/hooks/useSettings.ts
index f8f6aad6..d3d163fc 100644
--- a/app/src/lib/hooks/useSettings.ts
+++ b/app/src/lib/hooks/useSettings.ts
@@ -51,7 +51,9 @@ export function useCaptureSettings() {
if (
patch.stt_model !== undefined ||
patch.llm_model !== undefined ||
- patch.auto_refine !== undefined
+ patch.auto_refine !== undefined ||
+ patch.custom_llm_endpoint !== undefined ||
+ patch.custom_llm_model !== undefined
) {
queryClient.invalidateQueries({ queryKey: ['capture-readiness'] });
}
diff --git a/backend/backends/openai_compat_backend.py b/backend/backends/openai_compat_backend.py
index ce52891f..822eac18 100644
--- a/backend/backends/openai_compat_backend.py
+++ b/backend/backends/openai_compat_backend.py
@@ -117,7 +117,7 @@ class OpenAICompatLLMBackend:
headers["Authorization"] = f"Bearer {self.api_key}"
url = f"{self.endpoint}/chat/completions"
- async with httpx.AsyncClient(timeout=self.timeout) as client:
+ async with httpx.AsyncClient(timeout=self.timeout, follow_redirects=True) as client:
try:
response = await client.post(url, headers=headers, json=payload)
response.raise_for_status()
diff --git a/backend/services/llm.py b/backend/services/llm.py
index e89c9f8c..ff9cfc9f 100644
--- a/backend/services/llm.py
+++ b/backend/services/llm.py
@@ -2,7 +2,7 @@
LLM inference module - delegates to backend abstraction layer.
"""
-from ..backends import get_llm_backend, LLMBackend
+from ..backends import LLMBackend, get_llm_backend
def get_llm_model() -> LLMBackend:
@@ -10,6 +10,20 @@ def get_llm_model() -> LLMBackend:
return get_llm_backend()
+def resolve_model_size(backend: LLMBackend, requested: str | None) -> str:
+ """Model label to record for a call on ``backend``.
+
+ The remote OpenAI-compatible backend ignores Qwen size requests and
+ reports its own model name, so the label persisted on captures and
+ returned to clients must come from the backend, not the request.
+ """
+ from ..backends.openai_compat_backend import OpenAICompatLLMBackend
+
+ if isinstance(backend, OpenAICompatLLMBackend):
+ return backend.model_size
+ return requested or backend.model_size
+
+
def unload_llm_model() -> None:
"""Unload LLM model to free memory."""
get_llm_backend().unload_model()
diff --git a/backend/services/personality.py b/backend/services/personality.py
index a9027847..6449dde9 100644
--- a/backend/services/personality.py
+++ b/backend/services/personality.py
@@ -82,7 +82,7 @@ async def compose_as_profile(
"""
text = _require_personality(personality)
backend = llm_service.get_llm_model()
- resolved_size = model_size or backend.model_size
+ resolved_size = llm_service.resolve_model_size(backend, model_size)
system_prompt = _build_system_prompt(text, _COMPOSE_TASK)
output = await backend.generate(
@@ -107,7 +107,7 @@ async def rewrite_as_profile(
raise ValueError("Rewrite needs non-empty text to restate.")
backend = llm_service.get_llm_model()
- resolved_size = model_size or backend.model_size
+ resolved_size = llm_service.resolve_model_size(backend, model_size)
system_prompt = _build_system_prompt(character, _REWRITE_TASK)
output = await backend.generate(
diff --git a/backend/services/refinement.py b/backend/services/refinement.py
index b23d2192..ec033a39 100644
--- a/backend/services/refinement.py
+++ b/backend/services/refinement.py
@@ -277,7 +277,7 @@ async def refine_transcript(
produced the refinement.
"""
backend = llm_service.get_llm_model()
- resolved_size = model_size or backend.model_size
+ resolved_size = llm_service.resolve_model_size(backend, model_size)
# Pre-process before the LLM sees the text — the model shouldn't have
# to reason about obvious STT garbage (see ``collapse_repetitive_artifacts``).
diff --git a/backend/tests/test_openai_compat_backend.py b/backend/tests/test_openai_compat_backend.py
index 225681dc..cbb887f7 100644
--- a/backend/tests/test_openai_compat_backend.py
+++ b/backend/tests/test_openai_compat_backend.py
@@ -50,6 +50,9 @@ class FakeOpenAIServer:
status, payload = server.script(self.path, self.headers, body)
data = payload if isinstance(payload, bytes) else json.dumps(payload).encode()
self.send_response(status)
+ if status in (301, 302, 307, 308):
+ self.send_header("Location", payload["location"])
+ data = b""
self.send_header("Content-Type", "application/json")
self.send_header("Content-Length", str(len(data)))
self.end_headers()
@@ -318,3 +321,30 @@ def test_routes_round_trip_through_fake_server(tmp_path):
assert not isinstance(get_llm_backend(), OpenAICompatLLMBackend)
readiness = client.get("/capture/readiness").json()
assert readiness["llm"]["display_name"] != "remote-model"
+
+
+async def test_generate_follows_redirects():
+ def script(path, _headers, _body):
+ if path == "/v1/chat/completions":
+ return 307, {"location": "/v2/chat/completions"}
+ return _chat_reply("moved")
+
+ with FakeOpenAIServer(script) as server:
+ backend = OpenAICompatLLMBackend(endpoint=server.base_url, model="m")
+ assert await backend.generate(prompt="x") == "moved"
+ assert [r["path"] for r in server.requests] == ["/v1/chat/completions", "/v2/chat/completions"]
+
+
+async def test_refinement_and_personality_record_remote_model_label():
+ from backend.services import personality, refinement
+
+ with FakeOpenAIServer(lambda *_: _chat_reply("Refined.")) as server:
+ set_llm_config(server.base_url, "remote-model", None)
+ text, size = await refinement.refine_transcript(
+ "um hello", refinement.RefinementFlags(), model_size="0.6B"
+ )
+ assert (text, size) == ("Refined.", "remote-model")
+ result = await personality.rewrite_as_profile("A pirate.", "hello", model_size="0.6B")
+ assert result.model_size == "remote-model"