Files
jobboard/config.example.yaml
T
WolfandClaude Sonnet b560cdce2a Switch the embed model to multilingual-e5-small for the European corpus
Measured on the live roles against the "Copenhagen engineering" agent's
intent: the English-only all-MiniLM-L6-v2 separates engineering from
non-engineering roles with an AUC of 0.68 overall but only 0.74 restricted
to Danish-country roles (a Danish engineering role beats the rest's
non-engineering median only 68% of the time). multilingual-e5-small reaches
0.70 overall and 1.00 on Danish-country roles, for about 3x the CPU time.
paraphrase-multilingual-MiniLM-L12-v2 measured worse than the English-only
baseline (AUC 0.60) and was dropped.

e5 needs "query: "/"passage: " prefixes; app/jobs/embed.py adds them by
matching the model name so a model change can't leave a stale prefix
setting behind. Added tests for the prefixing and for the two invariants
the model change relies on: embed drops the old model's vectors before
re-embedding, and an agent's cached intent vector is keyed by model so a
model change is a cache miss, never a stale-dimension read.

Co-Authored-By: Claude Sonnet <noreply@anthropic.com>
2026-09-29 13:59:47 +02:00

135 lines
6.4 KiB
YAML

# roledesk configuration. The app reads --config, else $APP_CONFIG, else
# config.yaml beside the checkout (committed: these defaults plus the starter
# agents). Every value below is the default, so a copy with nothing changed
# behaves exactly like no file at all. An unknown key is an error, not a
# shrug; see app/config.py.
server:
# 0.0.0.0 in a container -- the proxy is what is in front of it.
host: 127.0.0.1
port: 8790
# The SQLite file: roles, companies, subscriptions, applications, agents and
# scores, the job queue and its schedules, the notice log, and auth's
# sessions. Not shared with newsdesk.
database: roledesk.db
# What a clock on screen means, and what `daily at 06:00` is 06:00 in.
timezone: Europe/Berlin
log_level: INFO
# One password on a public hostname; see auth/README.md and
# docs/password-protection.md. The password is never in this file.
auth:
enabled: auto # auto: open on loopback, locked on anything else
session_days: 30
secure_cookie: auto # always, behind a TLS-terminating proxy
rp_id: "" # roledesk.example.com -- passkeys are bound to it
rp_name: app # what the authenticator shows you
origin: "" # https://roledesk.example.com
password_env: APP_PASSWORD
cookie: app_session # unique per app sharing a hostname
# Any .py file in these directories defining a desk `Source` subclass is a
# plugin. A plugin imports `desk` only.
plugin_paths:
- plugins
# Workers per resource class. `http` is wide because a fetch is mostly waiting.
workers:
llm: 4
http: 8
cpu: 1
# Who answers when a job needs a model. See desk/tools/agent.py; needs the
# `agent` extra.
agent:
backend: auto # auto | api | claude_code | off
# Title and description in, gate columns out (app/jobs/extract.py). Off by
# default: on the API backend every batch of 20 roles is a paid call. On, a
# pull reads only the best roles of every enabled agent, as deep and as low
# as that agent's own limits say (see agents below), after a first score; a
# second score then applies the gates to what was filled.
extract:
enabled: false
model: claude-haiku-4-5-20251001
# About 60 words on a role: what it is, the stack, where, the pay if stated
# (app/jobs/summarise.py). Off by default. On, each pull summarises the best
# roles of every enabled agent that have none yet, by that agent's limits, so
# the cost stays bounded however many roles are open. The Summarise button on
# a role's page asks for that one role either way.
summarise:
enabled: false
model: claude-haiku-4-5-20251001
# A role and each agent's intent as vectors, for the fit term of a score
# (app/jobs/embed.py). Needs the `embeddings` extra; without it `embed` posts
# a notice and scores leave fit out. Changing the model re-embeds.
# multilingual-e5-small, not English-only MiniLM, because the corpus is now
# European: measured against the "Copenhagen engineering" agent's intent, it
# separates engineering from non-engineering roles better, especially among
# Danish-country roles (see docs/decisions.md), for about 3x the CPU time. An
# e5 model's "query: " / "passage: " prefixes are added automatically, keyed
# off this name.
embed:
model: intfloat/multilingual-e5-small
per_run: 50000
# One block per plugin, keyed by the plugin's name. The ats and search plugins
# read their boards and searches from the subscriptions table, so they need
# none, except the key an adzuna search needs (or ADZUNA_APP_ID and
# ADZUNA_APP_KEY in the environment, which keeps it out of this file).
sources: {}
# search:
# adzuna: {app_id: "...", app_key: "..."}
# Starter search agents, imported once into an empty `agents` table. Each gate
# is {field, values, mode: require|any|prefer, weight}; the any gates are one
# group a role must match one of ("in Denmark, or remote"), and a gate rejects
# only a role whose column is filled, since extract fills the rest for the
# roles that rank. Fields are place, country,
# remote, seniority, function, salary_min, currency, permit, languages and
# company. weights default to {fit: 0.6, recency: 0.2, affinity: 0.2, taste:
# 0.2}; taste is how close a role is to the ones the agent liked minus the
# ones it passed. A country is an ISO code, filled from the board's location
# line; EU is a role open anywhere in Europe.
# limits is how deep into this agent's ranking the two model stages read,
# each a top N (the backstop) and a minimum score (the real stop): extract
# reads the best extract_top roles at extract_min or more, summarise the best
# summarise_top at summarise_min or more. A fresh role with the worst fit
# scores about 0.27 on the default weights and a good one 0.6 or more. The
# defaults are {extract_top: 500, extract_min: 0.3, summarise_top: 30,
# summarise_min: 0.5}; a wide monitoring search raises them, a narrow hunt
# lowers them.
# The Search agents sub-tab lists and edits them; the order menu on Roles
# picks one by name. Two more, beside the starter in config.yaml:
agents: []
# - name: Remote from Europe
# intent: >-
# Staff or lead engineer role, fully remote from Europe. Platform and
# infrastructure work, distributed systems, developer tooling, technical
# leadership across teams.
# gates:
# - {field: remote, values: [remote], mode: require}
# - {field: country, values: [AT, BE, BG, HR, CY, CZ, DK, EE, FI, FR, DE, GR, HU, IE, IT,
# LV, LT, LU, MT, NL, PL, PT, RO, SK, SI, ES, SE, IS, LI, NO,
# GB, CH, EU], mode: require}
# limits: {extract_top: 1000, extract_min: 0.3, summarise_top: 50, summarise_min: 0.5}
# - name: Europe hybrid
# intent: >-
# Senior backend or platform engineer in a European tech hub, hybrid or
# remote-first. Python and Go, distributed systems, cloud infrastructure.
# gates:
# - {field: country, values: [AT, BE, BG, HR, CY, CZ, DK, EE, FI, FR, DE, GR, HU, IE, IT,
# LV, LT, LU, MT, NL, PL, PT, RO, SK, SI, ES, SE, IS, LI, NO,
# GB, CH, EU], mode: require}
# - {field: remote, values: [hybrid, remote], mode: prefer, weight: 0.05}
# Starter searches, imported once while there are no `search` subscriptions.
# provider is jobindex (Denmark, no key: q, area, pages) or adzuna (the UK and
# most of the EU, a free key: q, country, where, pages). See plugins/search.py.
searches: []