perf(llm): short one-shot calls say so to Strata

Every call that no later request continues (call_llm_direct,
call_llm_streaming, the warm-up and the v1 intent judge) now sends
"strata_checkpoint": false on the OpenAI-compatible path. Strata's
Qwen3.6 engine then gives that call's session away first, so with two or
more requests at once a reply's tool turn keeps its session through the
short calls made between its steps. The reply loop's chat_with_messages
is not marked (its next step continues it); the Ollama API gets no new
field. With one request at once nothing changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
javis-bot
2026-11-17 07:06:25 +09:00
parent 07615ee7a1
commit 8c61d7a821
4 changed files with 50 additions and 4 deletions

View File

@@ -2,7 +2,7 @@
Every distinct LLM call in Jarvis, what feeds it, what consumes it, and how it is gated. This is the reference for optimising the app's main bottleneck (LLM latency). Keep it in sync with the code — see the note at the bottom.
**Transport (all contexts below).** `ollama_base_url` selects the API. A URL that ends in `/v1` (v2: Strata at `http://127.0.0.1:8080/v1`, which the app starts and writes with the model id Strata lists into the brain's `config/brain.json`, `app/src/strata.rs`; also llama.cpp or Ollama's own `/v1`) uses the OpenAI-compatible API: chat goes to `/chat/completions`, embeddings to `/embeddings`, and the warm-up is a one-token chat completion. Any other URL uses Ollama's `/api/chat`, `/api/generate` and `/api/embeddings`. The conversion lives in `src/jarvis/llm.py` only: callers keep sending Ollama-shaped messages and reading Ollama-shaped responses (`message.content`, `message.tool_calls` with dict arguments, `message.thinking` from `reasoning_content`). On the OpenAI path `num_ctx` and `keep_alive` are not sent (the server owns its context size and residency), `num_predict` becomes `max_tokens`, and the thinking switch becomes `chat_template_kwargs.enable_thinking`. The `num_ctx: 8192` limits listed below apply to the Ollama path. Ollama's own `/v1` ignores `num_ctx` and runs at its default context (2,048 tokens here), which cuts the system prompt; to test the OpenAI path on Ollama use a model tag created with a larger `num_ctx`. A server with no embedding model (Strata) answers `/embeddings` with an error and every embedding consumer fails open to its text path. In v2 the app also sets Strata's context per performance stage (`app/src/admin.spec.md`, "Performance stages"): the settings' `llm.context` (default 32,768 tokens) normally and 16,384 tokens with one request at a time while a game runs; while it switches the voice worker starts no LLM call (requests wait, at most 180 s, `voice_worker.spec.md` "LLM switches"), and while paused (a game with no room for the model) every brain call fails at once.
**Transport (all contexts below).** `ollama_base_url` selects the API. A URL that ends in `/v1` (v2: Strata at `http://127.0.0.1:8080/v1`, which the app starts and writes with the model id Strata lists into the brain's `config/brain.json`, `app/src/strata.rs`; also llama.cpp or Ollama's own `/v1`) uses the OpenAI-compatible API: chat goes to `/chat/completions`, embeddings to `/embeddings`, and the warm-up is a one-token chat completion. Any other URL uses Ollama's `/api/chat`, `/api/generate` and `/api/embeddings`. The conversion lives in `src/jarvis/llm.py` only: callers keep sending Ollama-shaped messages and reading Ollama-shaped responses (`message.content`, `message.tool_calls` with dict arguments, `message.thinking` from `reasoning_content`). On the OpenAI path `num_ctx` and `keep_alive` are not sent (the server owns its context size and residency), `num_predict` becomes `max_tokens`, and the thinking switch becomes `chat_template_kwargs.enable_thinking`. Every call that no later request continues (all `call_llm_direct` and `call_llm_streaming` calls, the warm-up and the v1 intent judge; not the reply loop's `chat_with_messages`, whose next step continues it) carries `"strata_checkpoint": false`: Strata's Qwen3.6 engine then gives that call's session away first, so with two or more requests at once a reply's tool turn keeps its session through the short calls made between its steps (with one request at once nothing changes, and Strata releases up to 0.1.59 ignore the field on this engine); other servers ignore it. The `num_ctx: 8192` limits listed below apply to the Ollama path. Ollama's own `/v1` ignores `num_ctx` and runs at its default context (2,048 tokens here), which cuts the system prompt; to test the OpenAI path on Ollama use a model tag created with a larger `num_ctx`. A server with no embedding model (Strata) answers `/embeddings` with an error and every embedding consumer fails open to its text path. In v2 the app also sets Strata's context per performance stage (`app/src/admin.spec.md`, "Performance stages"): the settings' `llm.context` (default 32,768 tokens) normally and 16,384 tokens with one request at a time while a game runs; while it switches the voice worker starts no LLM call (requests wait, at most 180 s, `voice_worker.spec.md` "LLM switches"), and while paused (a game with no room for the model) every brain call fails at once.
---

View File

@@ -36,7 +36,7 @@ def warm_up_ollama_model(base_url: str, model: str, timeout: float) -> bool:
if is_openai_compatible(base_url):
# The server loads its model at launch; one token checks it answers.
body = openai_chat_body(model, [{"role": "user", "content": "hi"}], stream=False,
thinking=False, max_tokens=1)
thinking=False, max_tokens=1, one_shot=True)
try:
response = post_chat_completion(base_url, body, timeout)
ok = response.status_code == 200
@@ -438,6 +438,7 @@ Examples:
self.config.model,
[{"role": "system", "content": system_prompt}, {"role": "user", "content": user_prompt}],
stream=False, thinking=self.config.thinking, max_tokens=200, temperature=0.0,
one_shot=True,
),
self.config.timeout_sec,
)

View File

@@ -110,7 +110,11 @@ def openai_chat_body(
max_tokens: Optional[int] = None,
temperature: Optional[float] = None,
tools: Optional[List[Dict[str, Any]]] = None,
one_shot: bool = False,
) -> Dict[str, Any]:
"""``one_shot``: no later request continues this one (a classification, an extraction, a
warm-up). Strata then gives its session away first, so a reply's tool turn keeps its own
session through the short calls made between its steps; other servers ignore the field."""
body: Dict[str, Any] = {
"model": chat_model,
"messages": _openai_messages(messages),
@@ -124,6 +128,8 @@ def openai_chat_body(
body["temperature"] = temperature
if tools:
body["tools"] = tools
if one_shot:
body["strata_checkpoint"] = False
return body
@@ -246,7 +252,8 @@ def call_llm_direct(base_url: str, chat_model: str, system_prompt: str, user_con
if is_openai_compatible(base_url):
caller = _caller_name()
body = openai_chat_body(chat_model, messages, stream=False, thinking=thinking, temperature=temperature)
body = openai_chat_body(chat_model, messages, stream=False, thinking=thinking, temperature=temperature,
one_shot=True)
try:
content = _openai_chat(base_url, body, timeout_sec, caller)["message"]["content"]
if content.strip():
@@ -386,7 +393,7 @@ def _openai_stream(
thinking: bool,
) -> Optional[str]:
"""Server-sent-events stream from ``/chat/completions``; answer text only (no reasoning)."""
body = openai_chat_body(chat_model, messages, stream=True, thinking=thinking)
body = openai_chat_body(chat_model, messages, stream=True, thinking=thinking, one_shot=True)
try:
with post_chat_completion(base_url, body, timeout_sec, stream=True) as resp:
resp.raise_for_status()

View File

@@ -231,3 +231,41 @@ def test_dictation_clean_up_uses_chat_completions():
out = _llm_clean_dictation("um hello there", V1, model="qwen")
assert out == "hello there"
assert post.call_args.args[0] == f"{V1}/chat/completions"
# One-shot calls (docs/llm_contexts.md, transport note): a call no later request continues says so
# ("strata_checkpoint": false), so Strata gives its session away first and a reply's tool turn keeps
# its own session through the short calls made between its steps.
def test_short_calls_are_marked_one_shot():
from jarvis.listening.intent_judge import IntentJudge, IntentJudgeConfig, warm_up_ollama_model
from jarvis.listening.transcript_buffer import TranscriptSegment
answer = json.dumps({"directed": True, "query": "q", "stop": False, "confidence": "high", "reasoning": "r"})
seg = TranscriptSegment(text="jarvis q", start_time=1.0, end_time=2.0)
calls = {
"direct": lambda: call_llm_direct(V1, "qwen", "s", "u"),
"streaming": lambda: call_llm_streaming(V1, "qwen", "s", "u"),
"warm-up": lambda: warm_up_ollama_model(V1, "qwen", timeout=5),
"intent judge": lambda: IntentJudge(IntentJudgeConfig(ollama_base_url=V1, model="qwen")).judge([seg], wake_timestamp=1.0),
}
for name, call in calls.items():
resp = _sse({"choices": [{"delta": {"content": "x"}}]}) if name == "streaming" else _response(_completion(answer))
with patch("jarvis.llm.requests.post", return_value=resp) as post:
call()
assert post.call_args.kwargs["json"].get("strata_checkpoint") is False, name
def test_the_reply_loop_is_not_marked_one_shot():
# its next step continues this one, so its session must be kept
with patch("jarvis.llm.requests.post", return_value=_response(_completion("ok"))) as post:
chat_with_messages(V1, "qwen", [{"role": "user", "content": "hi"}],
tools=[{"type": "function", "function": {"name": "getWeather"}}])
assert "strata_checkpoint" not in post.call_args.kwargs["json"]
def test_ollama_api_gets_no_one_shot_field():
with patch("jarvis.llm.requests.post", return_value=_response({"message": {"content": "ok"}})) as post:
call_llm_direct("http://127.0.0.1:11434", "m", "s", "u")
assert "strata_checkpoint" not in post.call_args.kwargs["json"]