config.yaml

# ARGUS configuration. Fork this repo and make it yours:
#   * set `title`, `digest.recipient`
#   * edit registry/sources.yaml (your queries, journals, feeds, keywords)
#   * edit `tags` (your workstreams) and prompts/research.md (your profile+rubric)
#   * pick a scoring provider below (several are free)
# MASTER SWITCH. `false` means scheduled sweeps do nothing at all — no
# collection, no scoring, no commit. A manual run of the pipeline workflow still
# works, so you can try one without editing this.
#
# It ships OFF in the template on purpose: an unconfigured copy would otherwise
# collect for the example field below, burn Actions minutes, and hand whatever
# it gathered to the next person who uses the template as their starting data.
# Turn it on once steps 1-3 of the README are done and your sources are yours.
enabled: false

title: "ARGUS"

window_days: 7          # collectors ignore items older than this
render_days: 14         # dashboard shows items from this window

# How items get scored. `llm` uses whichever provider you name below; `keyword`
# is the $0 fallback and needs no key at all. Either way, a missing API key
# degrades to keyword scoring rather than failing the run.
#
# provider presets (see argus/providers.py, README for the full table):
#   anthropic  ANTHROPIC_API_KEY   claude-haiku-4-5                    ~$1/$5 per Mtok
#   groq       GROQ_API_KEY        llama-3.3-70b-versatile             FREE tier
#   gemini     GEMINI_API_KEY      gemini-2.0-flash                    FREE tier
#   openrouter OPENROUTER_API_KEY  ...:free models                     FREE
#   openai     OPENAI_API_KEY      gpt-4o-mini
#   deepseek   DEEPSEEK_API_KEY    deepseek-chat
#   together   TOGETHER_API_KEY    Llama-3.3-70B-Instruct-Turbo
#   ollama     (no key)            llama3.1          local only — `make run`, not Actions
# Any other OpenAI-compatible endpoint works too: set provider to a name of
# your choosing plus `base_url` and `api_key_env`.
scoring:
  mode: llm             # llm | keyword
  provider: anthropic
  model: claude-haiku-4-5   # omit to take the provider's default
  # base_url:           # only for a provider not in the preset table
  # api_key_env:        # ditto — the env var holding the key
  batch_size: 20
  max_llm_items_per_run: 150

  # Peer-reviewed work is structurally handicapped and this corrects for it.
  # OpenAlex carries no abstract for 60-82% of recent articles from the large
  # commercial publishers (Elsevier deposits none to Crossref either), while
  # arXiv always has one — so a journal paper gets judged on its title while a
  # preprint gets judged on a full abstract. Measured here before this existed:
  # arXiv was 57% of everything above threshold.
  # Set both to 0 to rank purely on content, or raise `preprint` above 0 if you
  # would rather see preprints first.
  venue_weight:
    enabled: true
    peer_reviewed: 1        # published in a journal
    preprint: -1            # arXiv and other repositories
    no_abstract_offset: 1   # extra for a journal paper whose abstract is missing,
                            # since it was scored on a title alone

# Journal metadata frequently omits abstracts; try to recover them by DOI from
# Crossref. Free, no key. Gives up automatically after 10 consecutive misses —
# Elsevier deposits no abstracts, so for an Elsevier-heavy watchlist this costs
# ~10 requests once per run and then stops. Worth leaving on: it does work for
# MDPI, Springer, Wiley, IEEE and Taylor & Francis.
crossref_backfill:
  enabled: true
  limit: 60

prompt: prompts/research.md

# Reference documents. Drop a concept note, proposal, chapter outline or reading
# list into context/ and its text is appended to the scoring prompt, so items
# are judged against your actual project rather than a keyword list. Usually the
# highest-leverage tuning available — a two-page concept note says more about
# what you care about than any number of query strings.
# Plain text only (.md, .txt, .rst) and LLM scoring only; the keyword scorer
# does not read them.
context:
  enabled: true
  dir: context
  max_chars: 24000      # whole block is truncated here, so a big drop can't
                        # blow the context window or the bill

# Relevance floors, on the scorer's 0-10 scale. Nothing below `show_threshold`
# is ever displayed — an empty dashboard is a valid, expected outcome, not a
# failure, and nothing in ARGUS tries to hit a quota of items. Raise these to
# make the radar pickier; lower them to see more.
show_threshold: 6       # dashboard floor (and the digest floor, unless digest.min_score is set)
highlight_threshold: 8  # "Top Picks" band on the dashboard

# Workstreams — the tags the scorer assigns. Colors are Paul Tol *vibrant*.
# Replace these with your own field's themes (keep keys in sync with the prompt).
tags:
  planning:          {label: "Planning & modeling",   color: "#0077BB"}
  market-reform:     {label: "Market reform",         color: "#33BBEE"}
  industrial-power:  {label: "Industrial power",      color: "#CC3311"}
  regional:          {label: "Regional & cross-border", color: "#009988"}
  stakeholder:       {label: "Stakeholder intel",     color: "#EE7733"}
  methods:           {label: "Methods & tools",       color: "#EE3377"}

digest:
  enabled: true
  send_hour_utc: 0
  recipient: you@university.edu
  # min_score: 8        # email floor; defaults to show_threshold. Raise it to
                        # keep the daily email tighter than the dashboard.
                        # Zero qualifying items = no email that day.
  synthesis: none       # none | llm — a one-paragraph intro on top of the digest
  # By default synthesis reuses the scoring provider above. Override either:
  # synthesis_provider: anthropic
  synthesis_model: claude-sonnet-4-6

prune:
  keep_days: 180        # drop below-threshold items older than this
  keep_run_days: 30     # trim the per-source run log; only the last 3 entries
                        # per source are ever read (failure detection)

# Library-seeded discovery (opt-in): study your Zotero library, then surface the
# recent papers that CITE it or are by the authors you read most. digest.py pulls
# the top `top_n` into a "From your library" section. Only DOIs are sent to
# OpenAlex; the library itself never leaves the machine / is never committed.
library:
  enabled: true
  source: zotero_api        # zotero_api (auto, no manual export) | bib (local file)
  every_days: 7             # run the library sweep at most weekly (it's a shortlist)
  top_n: 30                 # size of the "From your library" digest shortlist

  # -- source: zotero_api -- fetch the library from zotero.org each run.
  # Create a READ-ONLY key at zotero.org/settings/keys and add it as the
  # ZOTERO_API_KEY Actions secret. Find your numeric id at zotero.org/settings/keys
  # (personal "userID") or on the group page URL (group id).
  zotero_library_type: user   # user (personal library) | group (shared lab library)
  zotero_library_id: "SET_ME"   # numeric Zotero userID (the API key lives in ZOTERO_API_KEY, never here)

  # -- source: bib -- alternatively, a local export at this path (gitignored).
  bib_path: library/zotero.bib

  discovery_days: 120       # lookback for citing papers / your authors' new work
  max_authors: 15           # follow this many of your most-read authors
  max_seed_works: 200       # cap library works used for citation discovery