Skip to content

Configuration

HireEx reads three classes of configuration:

  1. Environment variables — secrets and runtime toggles (.env).
  2. Candidate profile — your positioning, consumed by the scoring prompt (config/candidate_profile.toml).
  3. Prompts — tuned LLM instructions (prompts/*.md).

All three have .example templates tracked in the repo. The live files (.env, candidate_profile.toml, score_job.md, filter_job.md) are gitignored and stay on your disk only.

Environment variables

Defined via pydantic settings in core/config.py. The HIREEX_ prefix and double-underscore separator map nested fields (e.g. HIREEX_DATABASE__URL → Settings.database.url).

Variable Purpose Example
HIREEX_LLM__API_KEY OpenRouter (or adapter) key. sk-or-v1-...
HIREEX_LLM__MODEL Default scoring model slug. anthropic/claude-opus-4-7
HIREEX_LLM__FILTER_MODEL Cheap pre-filter model. anthropic/claude-haiku-4-5
HIREEX_LLM__HF_TOKEN Hugging Face token for embeddings. hf_...
HIREEX_DATABASE__URL Postgres URL. postgresql+asyncpg://...
HIREEX_INGESTION__LINKEDIN_USERNAME LinkedIn login. —
HIREEX_INGESTION__LINKEDIN_PASSWORD LinkedIn password. —
HIREEX_QDRANT__URL Vector store URL. http://localhost:6333
HIREEX_QDRANT__API_KEY Qdrant Cloud key (optional). —
REDIS_URL Redis connection. redis://redis:6379/0

Full reference: check core/config.py for the pydantic model. Any field there is an env var with the HIREEX_ prefix (root-level fields take no double-underscore).

Candidate profile

config/candidate_profile.toml drives the scoring prompt's "who is this candidate" block. The pydantic model (CandidateProfile in core/config.py) uses extra="allow", so new sections / keys can be added without code changes.

Key sections (see config/candidate_profile.example.toml for the full template):

  • [candidateProfile] — identity, summary, total_years, target_roles, and up to 3 proof_stack entries (portfolio projects the scorer may cite verbatim in its reasoning field).
  • [core] / [secondary] / [exploratory] — skill tiers. Scorer treats core as a production signal, secondary as a ramp, exploratory as out-of-scope.
  • [transferable_skills] — adjacent tools the scorer should not penalize you for missing.
  • [hard_no] — dealbreakers. Pairs with the scorer's DEALBREAKERS section.
  • [preferences] — preferred / avoid / favourite domains, work modes, reject phrases.
  • [preferences.positive_modifiers] — free-form dict the scoring prompt reads for whitelist / title-match boosts.
  • [skill_levels] — per-skill calibration the model can reference when ranking stack fit.
  • [experience_history] — compact role history with quantified highlights.
  • [constraints.geography] — preferred region, visa paths, on-site dealbreakers.
  • [constraints.salary] — floor, target_min, target_max, stretch in your target currencies.
  • [domain_focus] — primary domains you boost, avoid domains you reject.

The scoring prompt interpolates the serialized profile via {{ candidate_profile }}, so any field you add to the TOML is visible to the LLM on the next run.

Prompts

Two files, both reloaded per call (edit and re-run, no restart).

prompts/score_job.md

Full scoring rubric. Typical sections:

  • [ROLE] — one-liner on who the candidate is, followed by proof-stack.
  • [OUTPUT] — the exact JSON schema the model must emit.
  • [TIERS] — keyword buckets that trigger match-score adjustments.
  • [PRIORITY LADDER] — your actual preferences, applied first.
  • [DEALBREAKERS] — hard caps.
  • [MODIFIERS] — title / employer / geo boosts and penalties.
  • [INTERVIEW-BAR PENALTIES] — if you hate leetcode, put it here.
  • [GEO] — geography bonuses / penalties.
  • [SALARY_POTENTIAL] — comp bands calibrated to your profile.
  • [FEW-SHOT] — 6–10 labeled rows the model anchors to.
  • [INPUT] / [RULES] — template slots + final instructions.

Size budget: ≤ 6 KB. If it grows, delete the verbose parts the LLM doesn't actually need.

prompts/filter_job.md

Cheap pre-filter. Emits a minified JSON {match_score, confidence, reasoning}. Everything below the pre-filter threshold skips the expensive scorer.

Buckets are 5-tier (0.85+, 0.65–0.85, 0.40–0.65, 0.15–0.40, 0.00–0.15). Hard caps mirror the scorer dealbreakers so the two stages agree.

Size budget: ≤ 2 KB.

Live editing

api/routes/prompts.py exposes:

  • GET /prompt/score / PUT /prompt/score
  • GET /prompt/filter / PUT /prompt/filter

The dashboard Settings page binds to these. Changes take effect on the next scoring batch.

Where .example values become real

The apps always read the private filename (candidate_profile.toml, score_job.md, filter_job.md). The .example files are templates only — they're not loaded at runtime. If a private file is missing, the relevant route or script errors out with a clear message; there's no fallback to the example.