# Latest Developments — Source Configuration (ADR-0018)
#
# Two independent top-level sections:
#
#   sources:  flat list of fetch sources — read by src/pipeline/fetch.py
#   digest:   email digest settings     — read by src/digest/send.py
#
# Neither section references the other. Adding or removing a source never
# requires touching the digest section, and vice versa.
#
# Environment variables (set as GitHub Secrets):
#   EMAIL_SENDER, EMAIL_RECIPIENT, EMAIL_PROVIDER
#   EMAIL_PASSWORD or RESEND_API_KEY
#   YOUTUBE_API_KEY      — YouTube Data API v3
#   GEMINI_API_KEY       — Google AI Studio (free at aistudio.google.com)

# ---------------------------------------------------------------------------
# sources — flat list, one entry per source
#
# Every entry has these common fields:
#   type          rss | youtube | substack | hackernews | arxiv |
#                 huggingface | paperswithcode | operator_changelog |
#                 replicate | openreview | openrouter
#   name          human-readable label (also used as source_name on ProcessedItem)
#   source_class  primary | operator | practitioner | media | market
#   enabled       true | false  (default: true)
#
# Type-specific options follow inline (see per-type examples below).
# ---------------------------------------------------------------------------
sources:

  # ── RSS / Atom feeds ──────────────────────────────────────────────────────

  - type: rss
    name: "Anthropic Blog"
    url: "https://www.anthropic.com/rss.xml"
    source_class: operator
    enabled: true

  - type: rss
    name: "OpenAI Blog"
    url: "https://openai.com/blog/rss.xml"
    source_class: operator
    enabled: true

  - type: rss
    name: "Google AI Blog"
    url: "https://blog.google/technology/ai/rss/"
    source_class: operator
    enabled: true

  - type: rss
    name: "DeepMind Blog"
    url: "https://deepmind.google/blog/rss.xml"
    source_class: operator
    enabled: true

  - type: rss
    name: "Ethan Mollick — One Useful Thing"
    url: "https://www.oneusefulthing.org/feed"
    source_class: practitioner
    enabled: true

  - type: rss
    name: "Simon Willison's Weblog"
    url: "https://simonwillison.net/atom/everything/"
    source_class: practitioner
    enabled: true

  - type: rss
    name: "Together AI Blog"
    url: "https://www.together.ai/blog/rss.xml"
    source_class: operator
    enabled: true

  - type: rss
    name: "Lambda Blog"
    url: "https://lambda.ai/blog/rss.xml"
    source_class: operator
    enabled: true

  # SDK/library release feeds are disabled — they produce pure version-bump
  # items (e.g. "v0.8.3 released") with no substantive analysis content and
  # exhaust Gemini enrichment quota on noise.  Consider re-enabling individual
  # feeds only after adding a pre-enrichment filter that skips items whose
  # titles match a semver-only pattern (e.g. /^v?\d+\.\d+/).
  - type: rss
    name: "Groq SDK Releases"
    url: "https://github.com/groq/groq-python/releases.atom"
    source_class: operator
    enabled: false

  - type: rss
    name: "Cerebras SDK Releases"
    url: "https://github.com/Cerebras/cerebras-cloud-sdk-python/releases.atom"
    source_class: operator
    enabled: false

  - type: rss
    name: "Mistral Inference Releases"
    url: "https://github.com/mistralai/mistral-inference/releases.atom"
    source_class: operator
    enabled: false

  - type: rss
    name: "Ollama Releases"
    url: "https://github.com/ollama/ollama/releases.atom"
    source_class: practitioner
    enabled: false

  - type: rss
    name: "llama.cpp Releases"
    url: "https://github.com/ggerganov/llama.cpp/releases.atom"
    source_class: practitioner
    enabled: false

  - type: rss
    name: "LocalAI Releases"
    url: "https://github.com/mudler/LocalAI/releases.atom"
    source_class: practitioner
    enabled: false

  # ── YouTube channels ──────────────────────────────────────────────────────

  - type: youtube
    name: "Nate Jones"
    channel_id: "UC0C-17n9iuUQPylguM1d-lQ"
    source_class: practitioner
    enabled: true
    max_videos: 8

  - type: youtube
    name: "Wes Roth"
    channel_id: "UCqcbQf6yw5KzRoDDcZ_wBSw"
    source_class: practitioner
    enabled: true
    max_videos: 5

  - type: youtube
    name: "Matthew Berman"
    channel_id: "UCawZsQWqfGSbCI5yjkdVkTA"
    source_class: practitioner
    enabled: true
    max_videos: 5

  - type: youtube
    name: "The AI Daily Brief"
    channel_id: "UCKelCK4ZaO6HeEI1KQjqzWA"
    source_class: media
    enabled: true
    max_videos: 3

  # ── Substack newsletters ──────────────────────────────────────────────────

  - type: substack
    name: "Nate's Newsletter"
    slug: "natesnewsletter"
    source_class: media
    enabled: true

  # ── Singleton fetchers — one entry per fetcher ────────────────────────────

  - type: hackernews
    name: "Hacker News"
    source_class: practitioner
    enabled: true
    min_score: 100
    max_stories: 8
    keywords:
      - "LLM"
      - "large language model"
      - "Claude"
      - "GPT"
      - "Gemini"
      - "Llama"
      - "AI agent"
      - "transformer"
      - "RAG"
      - "fine-tuning"
      - "inference"
      - "benchmark"
      - "Anthropic"
      - "OpenAI"
      - "DeepMind"
      - "Mistral"
      - "Opencode"
      - "Claws"
      - "post training"
      - "copilot"
      - "cowork"

  - type: arxiv
    name: "arXiv"
    source_class: primary
    enabled: true
    categories:
      - cs.AI
      - cs.LG
      - cs.CL
      - cs.CV
      - cs.RO
    max_papers: 8

  - type: huggingface
    name: "HuggingFace Models"
    source_class: primary
    enabled: true
    max_models: 15
    min_downloads: 1000

  - type: paperswithcode
    name: "Papers with Code"
    source_class: primary
    enabled: true
    page_size: 15
    min_stars: 0

  - type: operator_changelog
    name: "Operator Changelogs"
    source_class: operator
    enabled: true
    feeds:
      - "https://www.anthropic.com/rss.xml"
      - "https://openai.com/blog/rss.xml"
      - "https://blog.google/technology/ai/rss/"
      - "https://deepmind.google/blog/rss.xml"

  - type: replicate
    name: "Replicate Models"
    source_class: operator
    enabled: false
    limit: 5

  - type: openreview
    name: "OpenReview Papers"
    source_class: primary
    enabled: true
    limit: 15
    venues:
      - "ICLR.cc/2025/Conference/-/Accepted"
      - "NeurIPS.cc/2025/Conference/-/Accepted"
      - "ICML.cc/2025/Conference/-/Accepted"

  - type: openrouter
    name: "OpenRouter Pricing"
    source_class: market
    enabled: false
    limit: 10

# ---------------------------------------------------------------------------
# digest — email digest settings
#
# Completely independent of the sources section. Filter predicates reference
# only ProcessedItem field names (Schema Contract B), never source names or types.
# ---------------------------------------------------------------------------
digest:
  subject: "Daily AI Digest — {date}"
  send_if_empty: false

  # Filters applied to ProcessedItem records before selecting digest items.
  # All three reference ProcessedItem field names — not source names or types.
  min_credibility: 0.3      # exclude items with credibility_score below this
  max_hype_risk: 0.8        # exclude items with hype_risk above this
  exclude_marketing: true   # exclude items where is_marketing is true

  # Gemini digest generation
  model: "gemini-2.5-flash"
  max_tokens: 2000
  max_items_per_source: 5

  prompt: |
    You are summarising content from AI/ML sources for a senior software engineer
    who follows the LLM space closely. Be concise and technical.

    For each item, write:
    Theme: [1–3 word label, e.g. "agentic RAG", "fine-tuning", "inference cost"]
    Summary: [2–3 sentences: what it is, why it matters, one concrete takeaway]

    Focus on:
    - New model releases and their key capabilities or benchmarks
    - Novel research techniques with practical implications
    - Important open-source releases (libraries, tools, datasets)
    - Significant industry news (funding, partnerships, policy)
    - Practical developer tutorials or insights worth knowing

    Skip:
    - Pure marketing announcements with no technical substance
    - Repetitive coverage of the same story
    - Beginner-level tutorials unless the technique is genuinely new

    After all items, add a ## TL;DR section: exactly 3 bullet points, each a
    single concise sentence naming the most impactful item and why it matters.
    Follow with one "Recurring theme: …" line.

    Finally, add a ## Item Themes section listing each item's URL and its theme
    label, one per line:
    - <url> | <theme label>

    And a ## Item Summaries section:
    - <url> | <summary>

# ---------------------------------------------------------------------------
# pipeline — processing pipeline settings
#
# gemini_model: fallback hint used only when the ListModels API is unreachable.
#   At runtime the pipeline calls ListModels, discovers all available Pro and
#   Flash models, and builds the cascade automatically in this priority order:
#
#     Tier 0 — Pro        (gemini-X.Y-pro)       highest capability, lowest RPD
#     Tier 1 — Flash      (gemini-X.Y-flash)      balanced
#     Tier 2 — Flash-Lite (gemini-X.Y-flash-lite) highest RPD, lightest
#
#   Within each tier: newest major.minor version first, stable before preview.
#   The cascade falls through to the next model only when daily quota (RPD) is
#   exhausted — never on rate-limit (RPM), which is handled by back-off.
#
#   This setting is prepended to the fallback cascade when the API is down, so
#   set it to your preferred starting model (default: gemini-2.5-pro).
#
#   Theme clustering and email digest always use gemini-2.5-flash regardless
#   of this setting (1 call/run; quality matters more than throughput there).
#
# gemini_rpm: minimum interval floor (60 / rpm seconds between calls).
#   The pipeline also adapts dynamically from x-ratelimit-remaining-requests
#   response headers — so this value only governs the fastest we'll ever go,
#   not a ceiling on how slow we'll back off when quota runs low.
#
# enrich_max_output_tokens: Gemini token budget per enrichment call; default 500.
#   With thinking_budget=0 (see src/pipeline/stages/enrich.py) the model no
#   longer burns output tokens on reasoning, so 500 is sufficient, but 800
#   provides headroom against longer SUMMARY fields.
# ---------------------------------------------------------------------------
pipeline:
  gemini_model: "gemini-2.5-pro"
  gemini_rpm: 15
  enrich_max_output_tokens: 800

# ---------------------------------------------------------------------------
# history — shared global section
# ---------------------------------------------------------------------------
history:
  enabled: true
  history_days: 7
  history_dir: history

# ---------------------------------------------------------------------------
# logging — shared global section
# ---------------------------------------------------------------------------
logging:
  level: "INFO"
  log_file: null