# Awesome-Embodied&MM system configuration.
# Everything domain/tuning-related lives here so behavior changes need no code edits.

provider:
  # OpenRouter is OpenAI-compatible. Key read from this env var.
  base_url: "https://openrouter.ai/api/v1"
  api_key_env: "OPENROUTER_API_KEY"
  # Sent as OpenRouter attribution headers (optional but recommended).
  http_referer: "https://github.com/wzii/Awesome_Embodied_MM"
  app_title: "Awesome-Embodied&MM"

models:
  # Tiered routing. Cheap = filtering/first-pass (high volume); mid = analysis; strong =
  # scoring. Chosen via a blind Opus-4.8 ranking of anonymized outputs on real papers (see
  # scripts/bench_models.py). Latency is NOT a selection criterion for these batch stages
  # (only interactive Q&A cares about latency, and its model is deferred):
  #   - cheap: deepseek-v4-flash — #1 on the scoring task (best skeptical N/A calibration,
  #     no fabrication) AND cheapest by ~50x; ideal for the high-volume relevance filter.
  #   - mid/strong: glm-5.1 — #1 overall in the blind ranking, low cost; its only weakness
  #     (27s latency) is irrelevant for batch work.
  # Excluded: deepseek-v4-pro (abdicated scoring -> all-N/A), glm-4.7-flash (4/6 parse
  # failures), gemini-3.1-flash-lite (fabricated scores).
  tiers:
    cheap: "deepseek/deepseek-v4-flash"
    mid: "z-ai/glm-5.1"
    strong: "z-ai/glm-5.1"
    xstrong: "openai/gpt-5.5"   # top tier for the most quality-critical stage (scoring)
  # Optional per-stage overrides (stage name -> tier). Falls back to the tier passed in code.
  stage_tiers:
    filter: cheap
    summarize: cheap
    analyze: mid
    extract: cheap      # deepseek-v4-flash: fast + JSON-reliable + faithful for table extraction
    score: xstrong      # flagship two-layer scoring -> gpt-5.5
    innovation: mid
    front_summary: strong
    qa: strong          # TODO: pick a low-latency Q&A model in Phase 7 (deferred)
  defaults:
    temperature: 0.2
    max_tokens: 4096
  # USD per 1,000,000 tokens, {input, output}. Used for cost tracking/logging.
  cost_per_million:
    "deepseek/deepseek-v4-flash": {input: 0.098, output: 0.197}
    "z-ai/glm-5.1": {input: 0.98, output: 3.08}
    "openai/gpt-5.5": {input: 5.00, output: 30.00}
    # alternates considered: gemini-3.5-flash {1.50, 9.00}, claude-sonnet-4.6 {3.00, 15.00}

constants:
  request_timeout: 90        # seconds per LLM call
  max_retries: 4
  retry_backoff: 3.0         # seconds, exponential base
  lookback_days: 60          # ignore papers older than this on fetch
  relevance_threshold: 0.5   # below this -> dropped
  analyze_cap: 40            # max papers fully analyzed per run (cost guard)

logging:
  level: "INFO"              # console level; debug file sink enabled via --debug/WAM_DEBUG
  log_dir: "logs"

email:
  feature_threshold: 6.0        # core papers with weighted_total >= this get a detailed card
  max_featured: 8               # cap on featured (detailed) papers
  max_grouped: 40               # cap on the grouped lower tier
  group_lower_by_direction: true  # group the lower tier by research direction (else flat list)

paths:
  db: "data/wam.db"
  pdf_cache: "cache/pdfs"
  index_dir: "data/index"

# Two-layer scoring weights. Top-4 WAM metrics weighted 2x; computed over non-N/A metrics.
scoring:
  general_weights:
    novelty: 1.0
    soundness: 1.0
    impact: 1.0
  wam_weights:
    generalist: 2.0
    inference_speed: 2.0
    specialist: 2.0
    inference_cost: 2.0
    trustworthiness: 1.0
    collaborative: 1.0
    controlled_generation: 1.0
    other: 1.0

# Source + search config is consumed in Phase 2 (fetch).
sources:
  arxiv:
    enabled: true
    categories: ["cs.RO", "cs.AI", "cs.LG", "cs.CV"]
    max_results: 300
  semantic_scholar:
    enabled: true
    api_key_env: "SEMANTIC_SCHOLAR_API_KEY"   # optional; higher rate limits if set
  papers_with_code:
    enabled: true
  news:
    enabled: true
    feeds:
      - "https://www.therobotreport.com/feed/"
      - "https://spectrum.ieee.org/feeds/topic/robotics.rss"
      - "https://huggingface.co/blog/feed.xml"

# Search keyword sets. The LLM filter makes the final core/adjacent/drop call,
# but these seed the source queries (broad net by design).
keywords:
  core:        # World Action Models proper
    - "world action model"
    - "action world model"
    - "embodied foundation model"
    - "robot foundation model"
    - "generalist robot policy"
  adjacent:    # transferable techniques — kept in the innovation track
    - "vision-language-action"
    - "world model"
    - "video generation"
    - "action-conditioned video"
    - "interactive world model"

qa:
  # Knowledge-base Q&A is long-context (no embeddings/vector DB): a compact pack of
  # paper summaries + scores + leaderboard + authors + trends is built from the DB and
  # passed to the qa-tier model. See stage_tiers.qa for the model.
  max_papers: 400            # cap papers in the pack (most-scored first) to bound context
  include_dropped: true      # include filtered-out papers (title-only) so the KB still covers them
  web_search: true           # also search the web (paper original text, related concepts)
  web_max_results: 5         # web results per query (OpenRouter web plugin; ~$0.02/query)

trends:
  sim_threshold: 0.62        # cosine edge threshold for the direction graph
  min_front_size: 3          # ignore clusters smaller than this
  window_days: 30            # momentum = recent window vs prior window

people:
  min_papers: 2              # influential if on >= N tracked papers
  min_citations: 2000        # (reserved) citation-based influence via Semantic Scholar
  max_per_run: 30            # cap influential authors processed per run (most prolific first)