Skip to content

Latest commit

 

History

History
90 lines (66 loc) · 4.81 KB

configuration.md

File metadata and controls

90 lines (66 loc) · 4.81 KB

Configuration

Configuration is split between environment variables (.env file), Django models (managed via interactive onboarding or Django Admin), and hardcoded defaults in linkedin/conf.py.

LLM Configuration (.env)

LLM settings are stored in .env (at assets/.env for Docker, or project root for local dev). Any OpenAI-compatible provider works. These are prompted during interactive onboarding if missing.

Variable Description Default
LLM_API_KEY API key for an OpenAI-compatible provider. (required)
AI_MODEL Model identifier for qualification, follow-up, and search keyword generation. (required)
LLM_API_BASE Base URL for the API endpoint. (none)

These can also be set as environment variables directly.

Campaign Settings (Django Model)

Campaign data is stored in the Campaign Django model (1:1 with common.Department), managed via Django Admin (/admin/) or created during interactive onboarding.

Field Type Description
product_docs text Product/service description. Used by LLM qualification, follow-up agent, and search keyword generation.
campaign_objective text Campaign goal. Used by LLM qualification, follow-up agent, and search keyword generation.
booking_link string URL included in follow-up messages when suggesting a meeting.
is_freemium boolean Whether this is a freemium campaign (uses KitQualifier instead of BayesianQualifier).
action_fraction float Target fraction of total connections for freemium campaigns.

Account Settings (Django Model)

Account data is stored in the LinkedInProfile Django model (1:1 with auth.User), managed via Django Admin or created during interactive onboarding.

Field Type Description Default
linkedin_username string LinkedIn login email. (required)
linkedin_password string LinkedIn password. (required)
active boolean Enable/disable this account. true
subscribe_newsletter boolean Receive OpenOutreach updates. true
connect_daily_limit integer Max connection requests per day. 20
connect_weekly_limit integer Max connection requests per week. 100
follow_up_daily_limit integer Max follow-up messages per day. 30
legal_accepted boolean Whether the user accepted the legal notice. false

Rate limiting is enforced by LinkedInProfile methods (can_execute(), record_action(), mark_exhausted()) backed by the ActionLog model, surviving daemon restarts.

GDPR Location Detection

On the first run, the daemon checks the logged-in user's LinkedIn country code against a static set of ISO-2 codes for jurisdictions with opt-in email marketing laws (EU/EEA, UK, Switzerland, Canada, Brazil, Australia, Japan, South Korea, New Zealand).

  • Non-GDPR location: subscribe_newsletter is auto-set to true for that account.
  • GDPR-protected location: the existing value is preserved (no override).
  • Unknown/empty location: defaults to GDPR-protected (errs on the side of caution).

This check runs once per account (a marker file in assets/cookies/ prevents re-runs).

Hardcoded Defaults (conf.py:CAMPAIGN_CONFIG)

Timing and ML defaults are hardcoded in linkedin/conf.py. These are not user-configurable.

Key Value Description
check_pending_recheck_after_hours 24 Base interval (hours) before first pending check. Doubles per profile via exponential backoff.
enrich_min_interval 1 Floor (seconds) between enrichment API calls during auto-discovery.
min_action_interval 120 Minimum seconds between major actions.
qualification_n_mc_samples 100 Monte Carlo samples for BALD computation.
min_ready_to_connect_prob 0.9 GP probability threshold for promoting QUALIFIED to READY_TO_CONNECT.
min_positive_pool_prob 0.20 P(f > 0.5) threshold for positive pool check in exploit mode.
embedding_model BAAI/bge-small-en-v1.5 FastEmbed model for 384-dim profile embeddings.
connect_delay_seconds 10 Delay between connect tasks.
connect_no_candidate_delay_seconds 300 Delay when candidate pool is empty.
check_pending_jitter_factor 0.2 Multiplicative jitter factor for backoff.
worker_poll_seconds 5 Sleep when task queue is empty.

Other constants: MIN_DELAY (5s) / MAX_DELAY (8s) for human-like wait timing.

Derived Paths

The system automatically generates these paths per account:

  • Cookie file: assets/cookies/<handle>.json (session persistence)
  • Model file: assets/models/campaign_<id>_model.joblib (per-campaign GP model)

See Templating for follow-up messaging configuration.