settings.py

import os

from core.paths import resource_dir

MODEL_NAME = "phi3"

OLLAMA_URL = "http://localhost:11434/api/generate"

# Ollama model-list endpoint (used by setup checks)
OLLAMA_TAGS_URL = "http://localhost:11434/api/tags"

# -------------------- #
# LLM runtime (Ollama)
# -------------------- #

# Context window. Prompts are a few hundred tokens; Ollama's default reserves far
# more KV cache, which is most of the RAM the model needs beyond its weights.
LLM_NUM_CTX = 2048

# Unload the model after 10 idle minutes so a busy machine gets its RAM back.
LLM_KEEP_ALIVE = "10m"

# A terse assistant should answer, not improvise.
LLM_TEMPERATURE = 0.2

# Warm-start skips preloading the model when less RAM than this (GiB) is free.
# phi3 loaded fine with 2.2 GiB free (Windows pages it in); 3.5 skipped the
# preload and the first question paid a 13.5s cold load.
MODEL_MIN_FREE_GB = 1.5

# Download size shown on the setup wizard's pull button.
MODEL_PULL_SIZE = "2.3 GB"

# After an Ollama failure, skip model calls for this long (seconds) and answer at
# once, instead of waiting out timeouts on every query.
LLM_DOWN_SECONDS = 30

# Wake-word sensitivity. 0.6 suited the old (degraded MME) capture path. With
# clean WASAPI capture + resampling (see openwakeword_listener), ambient sits
# near 0, but real "hey jarvis" utterances vary a lot by distance/articulation:
# measured peaks ranged 0.47 (quiet) to 0.97 (clear). On some voices/mics a quiet
# pass peaks ~0.5 but only crosses the bar for a SINGLE frame, so a 2-consecutive
# requirement silently dropped ~4 of 5 attempts (the "it can't hear me" bug). With
# ambient staying well under 0.3, a single frame over threshold is a safe trigger,
# so WAKE_CONSECUTIVE is 1. Raise it to 2 if you get false wakes; lower THRESHOLD
# if it still misses you.
WAKE_THRESHOLD = 0.3

WAKE_CONSECUTIVE = 1

SESSION_TIMEOUT = 20

VOICE_RATE = -25

VOICE_VOLUME = 1.0

# Speech capture (speech_recognition). adjust_for_ambient_noise can push the
# energy threshold very high in a noisy room (or right after TTS), forcing the
# user to almost shout. Cap it here so normal-volume speech is still picked up;
# dynamic_energy_threshold still adapts from this ceiling. Lower if Jarvis
# ignores you; raise if it triggers on background noise. (sr default is 300.)
MAX_ENERGY_THRESHOLD = 400

# Silence (seconds) that ends an utterance. 1.2 made every reply wait 1.2s
# after you stopped; 0.9 is snappier while still tolerating the gap between
# words (0.8 once clipped "how are you" to "how").
STT_PAUSE_SECONDS = 0.9

# Longest utterance (seconds). If end-of-speech is never detected (noise), this
# is the worst-case wait before Jarvis answers — it was 12.
STT_PHRASE_LIMIT = 8

MEMORY_SIMILARITY_THRESHOLD = 0.55

# Raised from 0.45: a weak match (e.g. a resume chunk scoring ~0.45 against a
# vague query like "how are you") was being injected into the prompt and the
# small model would confabulate around it. 0.6 keeps genuine document Q&A while
# rejecting tangential matches.
DOCUMENT_SIMILARITY_THRESHOLD = 0.6

VOSK_MODEL_PATH = os.path.join(
    str(resource_dir()),
    "models",
    "vosk",
    "vosk-model-small-en-us-0.15"
)

WAKE_WORD = "hey_jarvis"

WAKE_MODEL_PATH = os.path.join(
    str(resource_dir()),
    "models",
    "wake",
    "hey_jarvis_v0.1.onnx"
)

ONLINE_CHECK_HOST = "8.8.8.8"

# Second resolver probed when the first doesn't answer (some networks block one).
ONLINE_CHECK_FALLBACK_HOST = "1.1.1.1"

ONLINE_CHECK_PORT = 53

ONLINE_CHECK_TIMEOUT = 1.0

ONLINE_CACHE_TTL = 5.0

# Chosen at startup by mic auto-detect; None = PyAudio default.
# INPUT_DEVICE_INDEX feeds STT (speech_recognition): prefers an MME/DirectSound
# mic, which downmixes to mono cleanly and which Google STT transcribes well.
# WAKE_DEVICE_INDEX feeds the wake-word listener: prefers the same mic on WASAPI,
# whose clean shared-mode audio openWakeWord needs (MME's "communications"
# processing scores ~0). They are the same physical mic via different host APIs;
# both fall back to the PyAudio default when None.
INPUT_DEVICE_INDEX = None

WAKE_DEVICE_INDEX = None

# -------------------- #
# Online lookups (free, keyless)
# -------------------- #

# Wikipedia / DuckDuckGo / Open-Meteo answers when online. The question text is
# sent to those services; set False to keep every query on this PC.
ONLINE_LOOKUPS = True

# Seconds to wait for a lookup before giving up (then the LLM answers).
WEB_TIMEOUT = 4

# -------------------- #
# HUD (desktop overlay)
# -------------------- #

HUD_WS_HOST = "127.0.0.1"

HUD_WS_PORT = 8765

HUD_STATS_INTERVAL = 3.0

# Window size (logical px). Compact: title bar, orb, captions, input.
HUD_WIDTH = 380

HUD_HEIGHT = 360

# Painted before the page loads; matches the panel so there's no flash.
HUD_BACKGROUND = "#070d17"

# Windows 11 draws the rounded corners and this 1px outline natively (RGB).
HUD_BORDER_RGB = (46, 64, 84)

# Theme schedule (24h). Day -> cyan, evening -> gold, night -> frost.
HUD_THEME_DAY_START = 5

HUD_THEME_EVENING_START = 17

HUD_THEME_NIGHT_START = 21