▸ Quick model picker

Live · cached HF repos + Ollama daemon tags · changes apply on Save + AI server restart

scanning…

CHAT

— scanning HF cache + Ollama daemon — Save

IMAGE

— scanning — Save

VISION

— scanning — Save

▸ Browse recommended models + Add HF repo + Pull Ollama tag ↻ Rescan

▸ No chat model installed yet

Pick a recommended starter — we'll download it and set it as your default chat model in one click.

▸ Browse recommended starters

▸ Discord & routing core 9

DISCORD_TOKEN

Bot token from the Discord developer portal · required.

—

SEEKDEEP_ALLOWED_CHANNELS

Comma-separated channel IDs · empty = answer everywhere.

—

SEEKDEEP_BLOCKED_CHANNELS

Channels SeekDeep ignores even when mentioned.

—

SEEKDEEP_ADMIN_IDS

User IDs that get queue priority + persona/digest perms.

—

MAX_DISCORD_CHARS

Hard cap before SeekDeep splits an answer across messages.

—

CHAT_PROVIDER

Provider for chat completions.

nvidia-localopenaianthropic —

IMAGE_PROVIDER

Provider for image generation.

nvidia-localopenai —

VISION_PROVIDER

Provider for vision / image-understanding.

nvidia-localopenai —

MODEL_AUTO_FALLBACK

If a chat role's model fails to load, fall back to default_chat.

DIRTY

▸ Models & roles routing per task 8

LOCAL_CHAT_MODEL_ID

Default chat model — fallback when no role-specific model fits.

—

LOCAL_CHAT_QUALITY_MODEL_ID

Used when /ask is invoked with quality:high or for complex synthesis.

—

LOCAL_CHAT_REASONING_MODEL_ID

For math/logic queries · auto-routed by the intent classifier.

—

LOCAL_CHAT_FALLBACK_MODEL_ID

Soft fallback when the primary chat model can't load.

—

LOCAL_CHAT_LIGHTWEIGHT_MODEL_ID

For low-stakes routing under VRAM pressure.

—

LOCAL_IMAGE_MODEL_ID

SDXL base for txt2img / img2img.

—

LOCAL_VISION_MODEL_ID

Image-understanding · OCR mode + describe mode.

—

LOCAL_CHAT_QUANT

4bit fits 12-14B models on 24GB VRAM, ~1-2% quality drop.

4bit8bitnone DIRTY · requires restart

▸ VRAM & cache safety budgets 7

VRAM_SYSTEM_RESERVE_MB

Reserve for OS + background apps · keeps the desktop responsive under load.

—

VRAM_SAFETY_MARGIN_MB

Headroom for KV cache / allocator fragmentation.

—

MODEL_KEEP_MODE

task-lru = evict cold roles. always = keep all resident (needs 32GB+).

task-lrualways —

LOCAL_CHAT_KEEP_RESIDENT

Keep chat model in VRAM across task switches (zero reload cost).

—

LOCAL_IMAGE_KEEP_RESIDENT

Pin SDXL · useful if you batch image gens.

—

LOCAL_VISION_KEEP_RESIDENT

Pin vision model · rarely needed.

—

LOCAL_MODEL_CACHE_DIR

Where HF stores pulled weights.

—

▸ Paths & backends HF · Ollama · cache roots 4

▸ Where weights live on disk · how long Ollama keeps a model warm after the last request. Empty = use each tool's default.

HF_HOME

Hugging Face cache root. Override to put weights on a fast SSD or a separate volume.

—

OLLAMA_MODELS

Ollama models directory. Pulled tags live here.

—

OLLAMA_KEEP_ALIVE

Duration a model stays warm in Ollama after the last request. "0" = evict immediately · "5m" = keep warm 5 minutes · "-1" = forever.

—

LOCAL_AI_BASE_URL

FastAPI side-car URL. Loopback only — never expose this beyond the local box.

—

▸ Remote chat keys per-role API URL + key 10

⚠ Prompts for these roles leave the local machine. Keys are stored in .env in plaintext — same threat model as any local dev tool. Loopback only.

▸ Leave a row empty to keep the role on the local HF / Ollama path. Roles without a matching _BACKEND override stay local even if a key is set here.

LOCAL_CHAT_DEFAULT_API_URL

Base URL for the default_chat role when its backend is remote.

—

LOCAL_CHAT_DEFAULT_API_KEY

Bearer / x-api-key for the default_chat role.

—

LOCAL_CHAT_LIGHTWEIGHT_API_URL

Base URL for lightweight_chat — used for greetings, translations, trivial routing.

—

LOCAL_CHAT_LIGHTWEIGHT_API_KEY

Key for the lightweight_chat role.

—

LOCAL_CHAT_QUALITY_API_URL

Base URL for quality_text — synthesis-heavy queries. Anthropic uses https://api.anthropic.com/v1/messages.

—

LOCAL_CHAT_QUALITY_API_KEY

Key for quality_text.

—

LOCAL_CHAT_REASONING_API_URL

Base URL for reasoning_code. Groq / OpenRouter / DeepSeek all speak the openai-compat shape.

—

LOCAL_CHAT_REASONING_API_KEY

Key for reasoning_code.

—

LOCAL_CHAT_FALLBACK_API_URL

Base URL for the cold-failover role.

—

LOCAL_CHAT_FALLBACK_API_KEY

Key for fallback_chat.

—

▸ Memory conversation context 5

SEEKDEEP_MEMORY_SCOPE

user = per-user context · channel = shared (small private servers).

userchannel —

SEEKDEEP_MEMORY_MODE

rolling = sliding window · summarize = compact older turns.

rollingsummarize —

MAX_CONTEXT_MESSAGES

How many messages SeekDeep stores per scope.

—

MAX_CONTEXT_CHARS

Storage cap. ~12K tokens at 40K chars.

—

SEEKDEEP_MEMORY_RECENT_ENTRIES

Render cap · how many entries are actually sent to the model.

—

▸ Image generation SDXL pipeline 11

IMAGE_GUIDANCE_SCALE

Dreamshaper-XL sweet spot 6.5-8.0.

DIRTY

IMAGE_STEPS

28 is standard · /image quality:low|standard|high → 12/28/40.

—

IMAGE_SCHEDULER

dpmsolver++ (DPM++ 2M Karras) is tuned for Dreamshaper-XL.

dpmsolver++euler-addimdefault —

SEEKDEEP_IMAGE_PROMPT_DYNAMIC_REFINEMENT

Refine prompts via the pinned chat model before SDXL.

—

SEEKDEEP_IMAGE_PROMPT_DYNAMIC_MAX_WORDS

Cap on refined-prompt length in words.

—

SEEKDEEP_IMAGE_PROMPT_MAX_CHARS

Hard cap on the prompt sent to SDXL.

—

SEEKDEEP_UPSCALE_METHOD

lanczos is local/offline · realesrgan requires the optional model.

lanczosrealesrgan —

SEEKDEEP_UPSCALE_SHARPEN

Apply mild unsharp mask after upscale.

—

SEEKDEEP_UPSCALE_SHARPEN_RADIUS

Sharpen radius · higher = more pronounced edges.

—

SEEKDEEP_UPSCALE_SHARPEN_PERCENT

Sharpen intensity (0-200).

—

SEEKDEEP_UPSCALE_MAX_OUTPUT_PIXELS

Safety ceiling · prevents accidental 8k explosions.

—

▸ Chat generation sampling 4

CHAT_WEB_TEMPERATURE

Web-routed answers · lower = more coherent synthesis.

—

CHAT_TOP_K

Top-K sampling cap · 0 = disabled.

—

CHAT_WEB_MAX_NEW_TOKENS

Generation cap for web-backed answers (room for synthesis).

—

WEB_APPEND_SOURCES

Append "Sources:" line to web-routed answers.

—

▸ Feature flags SEEKDEEP_FEATURE_* 9

▸ The three reaction features (FORCE_REACT, EMOJI_VAULT, AUTO_REACT) default off so SeekDeep doesn't fight demonbot in shared servers.

IMG2IMG

Transform attached / replied images.

UNKNOWN

INSTRUCT_PIX2PIX

Edit-by-instruction · pulls extra model.

UNKNOWN

INPAINT

CLIPSeg-masked SDXL fill.

UNKNOWN

UPSCALE_REALESRGAN

Heavier upscale path · requires model.

UNKNOWN

NSFW_GATE

Hard refusal on NSFW prompts.

UNKNOWN

TTS_VOICE

Voice replies in voice channels.

UNKNOWN

EMOJI_VAULT

@SeekDeep emoji backup/import/count/list.

UNKNOWN

FORCE_REACT

Manual react via paginated emoji picker.

UNKNOWN

AUTO_REACT

Per-message rule scan · @SeekDeep reactrule admin.

UNKNOWN

▸ Integrations channel-targeted features 1

▸ Features that post to a specific channel. Save here writes .env; click ⟳ Reload .env in Launcher to apply.

Member join/leave log UNKNOWN

Posts a Member Joined / Member Left embed (avatar · mention · user tag + ID · account created · server member count) to the channel below. Set the channel ID in JOIN_LEAVE_CHANNEL_ID (.env). Requires the GuildMembers privileged intent — enable it at discord.com/developers/applications → your bot → Bot → Privileged Gateway Intents → Server Members.

JOIN_LEAVE_CHANNEL_ID UNKNOWN

post joins

SAVED

post leaves

SAVED

▸ Test post

▸ Personas tone · synced with Tweaks panel 4

▸ Single source of truth. The Tweaks panel cycles the same value live · this is where it persists to .env.

Clinical

terse, numbered, footnoted. minimal warmth. like a manpage with feelings.

Neurotic

second-guesses itself, asks clarifying questions, hedges. "are you sure that's what you want?"

Unsettling

too-knowing pauses. uses your name unprompted. ellipses everywhere…

Chaotic

CAPS LOCK. tangents. ENERGY. emoji approved. ships first asks later.

▸ Web search SearXNG 4

SEARXNG_BASE_URL

Local SearXNG instance · docker compose default.

—

WEB_AUTO_SEARCH

Let SeekDeep auto-decide when to call SearXNG.

—

WEB_SEARCH_STRICT_ROUTING

Require explicit web:always for force-search.

—

WEB_SEARCH_FAIL_OPEN

If SearXNG is down, answer from model knowledge instead of erroring.

—

▸ Admin offline mode · diagnostics 3

HF_HUB_OFFLINE

Lock Hugging Face to local cache only · run after first warmup.

01 —

TRANSFORMERS_OFFLINE

Lock Transformers to local cache · pairs with HF_HUB_OFFLINE.

01 —

MODEL_ROUTER_LOG

Log every routing decision to logs/.

—