config.example.yaml

# Heta configuration template
#
# Usage:
#   1. cp config.example.yaml config.yaml
#   2. Fill in your API keys in the [providers] section below
#   3. ./scripts/bootstrap.sh
#
# ┌─────────────────────────────────────────────────────────────────────────┐
# │  ZH TEMPLATE STYLE SETUP (no OpenAI / Gemini needed)                     │
# │  If you only have DashScope + SiliconFlow accounts, replace each        │
# │  <<: *openai  →  <<: *dashscope                                         │
# │  <<: *gemini  →  <<: *dashscope                                         │
# │  and change the model names as indicated in the comments below.         │
# └─────────────────────────────────────────────────────────────────────────┘


# ═══════════════════════════════════════════════════════════════════════════
# STEP 1 — Fill in your API keys here (everything else is pre-configured)
# ═══════════════════════════════════════════════════════════════════════════
providers:

  dashscope: &dashscope        # Alibaba Cloud DashScope — LLM + VLM + Embedding
    api_key: "YOUR_DASHSCOPE_API_KEY"
    base_url: "https://dashscope.aliyuncs.com/compatible-mode/v1/"

  siliconflow: &siliconflow    # SiliconFlow — Embedding
    api_key: "YOUR_SILICONFLOW_API_KEY"
    base_url: "https://api.siliconflow.cn/v1/"

  # ── Optional providers ───────────────────────────────────────────────────
  # Remove or leave empty if using only DashScope + SiliconFlow.

  openai: &openai              # OpenAI — MemoryKB LLM + Embedding
    api_key: "YOUR_OPENAI_API_KEY"
    base_url: "https://api.openai.com/v1"

  gemini: &gemini              # Google Gemini — HetaGen LLM
    api_key: "YOUR_GEMINI_API_KEY"
    base_url: "https://generativelanguage.googleapis.com/v1beta/openai/"


# ═══════════════════════════════════════════════════════════════════════════
# STEP 2 — Infrastructure
# Pre-configured for docker-compose. No changes needed.
# (Running without Docker? Change service names to 127.0.0.1)
# ═══════════════════════════════════════════════════════════════════════════
persistence:
  postgresql:
    host: postgres
    port: 5432
    user: postgres
    password: postgres
    database: postgres
    options: "-c search_path=public"

  milvus:
    host: milvus
    port: 19530
    url: "http://milvus:19530"

  neo4j:
    url: "bolt://neo4j:7687"
    username: "neo4j"
    password: "heta_password"
    database: "neo4j"


# ═══════════════════════════════════════════════════════════════════════════
# Module config — no changes needed for quick start
# ═══════════════════════════════════════════════════════════════════════════

heta:
  title: "Heta API"
  version: "0.1.0"
  fastapi:
    host: "0.0.0.0"
    port: 8000
    reload: false
    log_level: "info"


# ─────────────────────────────────────────────────────────────────────────
# Heta CLI — model used by CLI-only orchestration, such as answer judging.
# Default: Google Gemini
# ─────────────────────────────────────────────────────────────────────────
heta_cli:
  judge:
    <<: *gemini
    model: "gemini-3-flash-preview"
    max_retries: 3
    timeout: 60


# ─────────────────────────────────────────────────────────────────────────
hetadb:
  title: "HetaDB API"
  version: "0.1.0"
  workspace: "workspace"

  fastapi:
    host: "0.0.0.0"
    port: 8001
    reload: false
    log_level: "info"

  llm:
    <<: *dashscope
    model: "Qwen/Qwen3-32B"
    max_concurrent_requests: 10
    max_retries: 3
    timeout: 120

  vlm:
    <<: *dashscope
    model: "qwen3.5-plus"
    max_concurrent_requests: 10
    max_retries: 5
    timeout: 120

  embedding_api:
    <<: *siliconflow
    model: "BAAI/bge-m3"
    dim: 1024
    timeout: 30
    batch_size: 2000
    num_threads: 8
    max_file_size_bytes: 3221225472
    max_retries: 5
    retry_delay: 2

  # File storage — uses MinIO from docker-compose (S3-compatible)
  s3:
    ak: "minioadmin"
    sk: "minioadmin"
    endpoint: "http://minio:9000"            # internal address used by the backend (Docker network)
    public_endpoint: "http://localhost:9000" # address embedded in download URLs returned to clients
                                             # change to your server IP or domain for remote deployments
    bucket: "hetadb-raw-files"
    config:
      signature_version: "s3v4"
      retries:
        max_attempts: 3
        mode: "adaptive"
      read_timeout: 300
      connect_timeout: 60

  milvus:
    db_name: "hetadb"
    sentence_mode: false

  # Max concurrent dataset parse tasks.  Each task spawns graph.max_workers
  # LLM threads, so effective LLM concurrency = parse_max_workers × llm.max_concurrent_requests.
  parse_max_workers: 2

  query_defaults:
    top_k: 10
    threshold: 0.0
    similarity_weight: 1.5
    occur_weight: 1.0
    reranker_url: ""  # Optional: fill in your reranker service URL to enable reranking

  search_params:
    ef_multiplier: 10


# ─────────────────────────────────────────────────────────────────────────
hetamem:
  title: "HetaMem API"
  version: "0.1.0"

  fastapi:
    host: "0.0.0.0"
    port: 8003
    reload: false
    log_level: "info"

  memorykb:
    # Default: OpenAI
    # Domestic alternative — replace the block below with:
    #   llm:
    #     <<: *dashscope
    #     model: "qwen-plus"
    llm:
      <<: *openai
      model: "gpt-4o-mini-2024-07-18"

    # Default: OpenAI text-embedding-3-small (dim=1536)
    # Domestic alternative — replace the block below with:
    #   embedding:
    #     <<: *dashscope
    #     model: "text-embedding-v4"
    #     dim: 1536
    embedding:
      <<: *openai
      model: "text-embedding-3-small"
      dim: 1536

  memoryvg:
    # LLM: already using DashScope by default
    llm:
      provider: "openai"
      config:
        <<: *dashscope
        model: "qwen3-max"

    # Default: OpenAI text-embedding-3-large (dim=1024)
    # Domestic alternative — replace the block below with:
    #   embedder:
    #     provider: "openai"
    #     config:
    #       <<: *siliconflow
    #       model: "BAAI/bge-m3"
    #       embedding_dims: 1024
    embedder:
      provider: "openai"
      config:
        <<: *openai
        model: "text-embedding-3-large"
        embedding_dims: 1024

    vector_store:
      provider: "milvus"
      config:
        collection_name: "memoryvg"
        embedding_model_dims: 1024
        metric_type: "IP"
        db_name: "hetamem"

    graph_store:
      provider: "neo4j"

    version: "v1.1"


# ─────────────────────────────────────────────────────────────────────────
tasks:
  # Maximum number of tasks kept in memory across all modules.
  # When exceeded, oldest terminal-state tasks (completed/failed/cancelled) are evicted first.
  max_tasks: 500
  # Seconds before a terminal-state task is eligible for cleanup.
  ttl_seconds: 3600


# ─────────────────────────────────────────────────────────────────────────
hetagen:
  fastapi:
    host: "0.0.0.0"
    port: 8002
    reload: false
    log_level: "info"

  # Default: Google Gemini
  # Domestic alternative — replace the block below with:
  #   llm:
  #     <<: *dashscope
  #     model: "qwen3-32b"
  #     max_concurrent_requests: 10
  #     max_retries: 5
  #     timeout: 120
  llm:
    <<: *gemini
    model: "gemini-3-flash-preview"
    max_concurrent_requests: 10
    max_retries: 5
    timeout: 120

  vlm:
    <<: *siliconflow
    model: "Qwen/Qwen3-VL-32B-Instruct"
    max_concurrent_requests: 10
    max_retries: 5
    timeout: 120

  embedding_api:
    <<: *siliconflow
    model: "BAAI/bge-m3"
    dim: 1024
    timeout: 30

  milvus:
    db_name: "hetagen"
    collection_name: "top_100_embeddings"
    embedding_dim: 1024
    index_type: "IVF_FLAT"
    metric_type: "IP"
    nlist: 128
    batch_size: 100


# ─────────────────────────────────────────────────────────────────────────
hetawiki:
  # Wiki workspace root — relative to project root, or absolute path.
  # All raw uploads and generated wiki pages are stored here.
  workspace: "workspace"

  ingest:
    # Concurrent ingest workers. Each worker makes LLM API calls.
    # Increase only if your LLM API quota supports higher concurrency.
    max_workers: 4
    # Default ingest is single-page only. Reject overly long parsed documents
    # instead of silently truncating or auto-splitting them.
    max_input_chars: 80000
    # Timeout in seconds for document parsing (MinerU).
    # Large PDFs or scanned documents with OCR may take several minutes.
    parse_timeout: 300

  llm:
    # LLM used for wiki page generation.
    # Needs long-context support — documents can be tens of thousands of tokens.
    # Timeout is intentionally high: large documents take time.
    <<: *dashscope
    model: "Qwen/Qwen3-32B"
    timeout: 300
    max_retries: 2