DB GameAgent

DB GameAgent is a local game-database assistant. The current implementation targets Arknights data, story JSON, and image assets, then answers through a cloud or local LLM provider.

The app includes:

• FastAPI backend
• SQLite + FTS5 index
• React/Vite web UI
• chat history in browser localStorage
• local image lookup
• small user memory
• external wiki/web search
• configurable LLM providers

Examples

Supported LLM Providers

Configured from .env, selectable in the UI:

• BotHub
• OpenRouter
• x.ai
• OpenAI
• Gemini
• Local OpenAI-compatible API

The local provider can point to KoboldCPP, llama.cpp server, Text Generation WebUI, vLLM, LM Studio, or another server that exposes OpenAI-compatible /v1/chat/completions and optionally /v1/models.

Requirements

Install these first:

• Python 3.10+
• Node.js 20+ or 22+
• Git

Node.js is required because the UI is a Vite/React app.

Quick Start

1. Clone this repository.
2. Copy .env.example to .env.
3. Fill at least one model provider in .env.
4. Clone the Arknights data repositories:

clone_data_repos.bat

5. Build the SQLite index:

rebuild_index.bat

6. Start backend and UI:

start.bat

7. Open:

http://127.0.0.1:5173

To stop background processes:

stop.bat

Data Repositories

The large Arknights data folders are intentionally not committed. They are external repositories:

• ArknightsGamedata from https://github.com/ArknightsAssets/ArknightsGamedata
• ArknightsGameData_Zh_CN from https://github.com/Kengxxiao/ArknightsGameData.git
• ArknightsStoryJson from https://github.com/050644zf/ArknightsStoryJson
• Arknight-Images from https://github.com/Aceship/Arknight-Images

Use clone_data_repos.bat to clone them into the expected paths.

Environment

Core settings:

LLM_PROVIDER=bothub
LLM_TEMPERATURE=0.2
LLM_TIMEOUT_SECONDS=120
ENABLE_MODEL_TOOLS=true
AGENT_MAX_TOOL_CALLS=6
AGENT_MAX_TOOL_RESULT_CHARS=12000

BOTHUB_API_KEY=
BOTHUB_MODEL=
BOTHUB_BASE_URL=https://bothub.chat/api/v2/openai/v1

OpenRouter:

LLM_PROVIDER=openrouter
OPENROUTER_API_KEY=
OPENROUTER_MODEL=
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1

x.ai:

LLM_PROVIDER=xai
XAI_API_KEY=
XAI_MODEL=
XAI_BASE_URL=https://api.x.ai/v1

OpenAI:

LLM_PROVIDER=openai
OPENAI_API_KEY=
OPENAI_MODEL=
OPENAI_BASE_URL=https://api.openai.com/v1

Gemini:

LLM_PROVIDER=gemini
GEMINI_API_KEY=
GEMINI_MODEL=

Local OpenAI-compatible server:

LLM_PROVIDER=local
LOCAL_BASE_URL=http://127.0.0.1:8080/v1
LOCAL_MODEL=
LOCAL_API_KEY=

For local servers without auth, leave LOCAL_API_KEY empty.

UI Controls

The UI supports:

• Russian / English / Chinese / Japanese / Korean language switch
• provider dropdown
• model dropdown loaded from provider /models
• model search
• manual model override
• temperature
• source count
• context character limit
• chat history message limit
• model tool call toggle
• Arknights wiki search toggle
• Endfield wiki search toggle
• Brave web search toggle

The "Sources" slider controls retrieval results, not LLM sampler top_k.

External Search

Wiki search:

• prts.wiki
• arknights.wiki.gg
• optional endfield.wiki.gg

Brave Search API can be enabled with:

WEB_SEARCH_ENABLED=true
BRAVE_SEARCH_API_KEY=

Endfield wiki search is off by default to avoid mixing Arknights and Arknights: Endfield context:

ENDFIELD_WIKI_SEARCH_ENABLED=true

Indexing

Run:

rebuild_index.bat

The index is stored in:

data/arknights_agent.sqlite

SQLite is used because this is a local, single-user app. Postgres can be added later for multi-user hosting or vector search.

Development

Backend:

.venv\Scripts\activate.bat
uvicorn backend.app.main:app --host 127.0.0.1 --port 8017

Frontend:

cd frontend
npm install
npm run dev

Production-style frontend build:

cd frontend
npm run build

Notes

• .env, data folders, SQLite files, node_modules, and local llama.cpp builds are ignored by git.
• Model keys are never stored in the UI.
• Gemini currently uses retrieval-first prompting without model tool calls.
• OpenAI-compatible providers can use model tool calls when enabled. Tool calls are locally schema-validated, permission-checked, bounded by a per-run tool-call budget, and returned to the model as structured observations.