Say Search agents, Viewing as and the want feed score in the config comments, add the cv calibration band and shorten the comments that restated the ADR

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
WolfandClaude Fable 5.1 committed 2026-10-05 09:32:43 +02:00
1 parent b64d77d224
commit c3bc2cd7f3
2 files changed
+62 -69

No files matched your search

+58 -65
View File
@@ -51,11 +51,10 @@ agent:
max_tokens: 8192 # the answer's cap per call
cli_path: "" # where the `claude` binary is, if not on PATH
# 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 cut of every enabled agent (its limits, see agents
# below), after a first score; a second score then applies the fetch gate and
# the cut to what was filled. `enabled` is only the
# Title and description in, gate columns out: location, remote area, seniority,
# languages and deadline (app/jobs/extract.py). Off by default: on the API
# backend every batch of 20 roles is a paid call. On, a pull reads the cut of
# every enabled agent (its limits, see agents below). `enabled` is only the
# default for a fresh database: the switch under Settings, Sources, Runs (or
# `app stages extract on|off`) overrides it and needs no restart.
extract:
@@ -66,12 +65,11 @@ extract:
thinking: off # off | adaptive
# 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 top
# of every enabled agent's feed that has none yet, the best summarise_top (its
# limits) by the want feed score and the best summarise_top by the match
# order (CV match, then CV closeness), and every role liked or saved, 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. `enabled` is only the
# (app/jobs/summarise.py). Off by default. On, each pull summarises the
# summarise_top (an agent's limits) best roles of each enabled agent's two
# orders, What I'm looking for and What fits me, and every role liked or
# saved, when they have no summary yet. The Summarise button on a role's page
# asks for that one role either way. `enabled` is only the
# default for a fresh database: the switch under Settings, Sources, Runs (or
# `app stages summarise on|off`) overrides it and needs no restart.
summarise:
@@ -84,12 +82,12 @@ summarise:
# Only extract reads it.
# Each search agent's cut read against its CV: skills, experience and career 0
# to 10, a total 0 to 100, and 1 to 3 strengths and gaps (app/jobs/judge.py),
# for the CV match order on Roles. Off by default. On, each pull judges the cut of
# every enabled agent with a CV and a prompt, and only the roles not judged
# yet or judged from another CV, prompt, like, avoid or must text, or role
# text. The CV rides in every call, so batches are
# small. `enabled` is only the default for a fresh database: the switch under
# to 10, and 1 to 3 strengths and gaps (app/jobs/judge.py); the CV match, which
# orders What fits me on Roles, weights the three. Off by default. On, each pull
# judges the cut of every enabled agent with a CV and a prompt, once per change
# of its CV, prompts, likes, dislikes or must text, or of the role's text. The
# CV rides in every call, so batches are small. `enabled` is only the default
# for a fresh database: the switch under
# Settings, Sources, Runs (or `app stages judge on|off`) overrides it and needs
# no restart. on_ingest is accepted and ignored, as for summarise.
judge:
@@ -110,25 +108,23 @@ fetch_text:
per_run: 1000 # reads per run
retry_days: 7 # days before a page that gave no text is read again
# A role and each agent's prompts 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.
# calibration maps a model's cosines to 0..1 levels, as [lo, hi] for fit and
# for taste: lo is 0, hi is 1, clipped outside. multilingual-e5-small has
# built-in constants (app/config.py, CALIBRATION); an entry here replaces
# them for that model. A model with none is min-max scaled per run instead,
# so its fit is a rank, not a level, and the log says so.
# A role, each agent's prompts and its CV as vectors, for prompt closeness and
# CV closeness (app/jobs/embed.py). Needs the `embeddings` extra; without it
# `embed` posts a notice and scores leave closeness out. Changing the model
# re-embeds. multilingual-e5-small, not English-only MiniLM, because the corpus
# is European (see docs/decisions.md). An e5 model's "query: " / "passage: "
# prefixes are added automatically, keyed off this name.
# calibration maps a model's cosines to 0..1 levels, as [lo, hi] for fit (prompt
# closeness), taste (job like) and cv (CV closeness, fit's band when left out):
# lo is 0, hi is 1, clipped outside. multilingual-e5-small has built-in
# constants (app/config.py, CALIBRATION); an entry here replaces them for that
# model. A model with none is min-max scaled per run instead, so its levels are
# ranks, and the log says so.
embed:
model: intfloat/multilingual-e5-small
per_run: 50000
calibration: {}
# intfloat/multilingual-e5-small: {fit: [0.75, 0.88], taste: [-0.14, 0.19]}
# intfloat/multilingual-e5-small: {fit: [0.75, 0.88], taste: [-0.14, 0.19], cv: [0.77, 0.90]}
# 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
@@ -141,39 +137,36 @@ sources: {}
# search:
# adzuna: {app_id: "...", app_key: "..."}
# Starter search agents, imported once into an empty `agents` table. The gates
# are the agent's fetch gate, each {field, values, mode: require|any|exclude}:
# a require gate is the region, the any gates are one group a role must match
# one of ("in Denmark, or remote"), the place part, and an exclude gate leaves
# out the roles that match one of its values (title: the titles to leave
# out). A gate rejects only a role whose column is filled, since extract fills
# the rest for the roles in the cut. A company set to All locations skips the
# any group, one set to All roles the exclude gates on title. Fields are
# place, country, remote, remote_from, seniority, function, salary_min,
# currency, permit, languages, company and title. weights are the final
# score's, the feed's order, default {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 the agent's cut, the roles that get fetch text, extract and
# judge: the best cut_fit_top by fit level at cut_fit_min or more, then the
# best cut_cv_top by CV closeness at cut_cv_min or more that the first list
# leaves out, then every role liked or saved. Both levels are calibrated, so
# a minimum means the same in every run, and a liked company's roles sit
# about 0.05 higher, a favourite's about 0.1. summarise_top is how deep into
# each order, want and match, summarise reads. The defaults are
# {cut_fit_top: 500, cut_fit_min: 0.6, cut_cv_top: 150, cut_cv_min: 0.6,
# summarise_top: 30}. For a broad agent the count stops a list, for a narrow
# one the minimum. An agent without a prompt that has text has no fit, so
# nothing of it is in the cut by fit.
# prompts is a list of {name, text}, each what the agent looks for; a role's
# fit is its best prompt's, so one agent can hunt several kinds of role
# without blending them into one paragraph. Votes, gates and limits stay the
# agent's, shared by its prompts. intent, as in config.yaml, is one prompt
# named after the agent.
# enabled (default true) leaves an agent out of the scores and the pull when
# false. 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:
# Starter search agents, imported once into an empty `agents` table. The Search
# agents page lists and edits them afterwards, and the Viewing as select on
# Roles picks one by name. An agent has:
# - name, and enabled (default true; false leaves it out of the scores and the
# pull).
# - prompts, a list of {name, text}, each what the agent looks for; a role is
# ranked by the prompt it is closest to. intent, as in config.yaml, is one
# prompt named after the agent.
# - gates, its fetch gate, each {field, values, mode: require|any|exclude}: a
# require gate is the region (the world), the any gates are one group a role
# must match one of ("in Denmark, or remote"; the location), and an exclude
# gate leaves out the roles that match one of its values (title: the titles
# to leave out). A gate rejects only a role whose column is filled, since
# extract fills the rest for the roles in the cut. A company set to All
# locations skips the any group, one set to All roles the exclude gates on
# title. Fields are place, country, remote, remote_from, seniority,
# function, salary_min, currency, permit, languages, company and title. A
# country is an ISO code, filled from the board's location line; EU is a
# role open anywhere in Europe.
# - weights, the want feed score's, which orders What I'm looking for: {fit:
# 0.6, recency: 0.2, affinity: 0.2, taste: 0.2}, that is prompt closeness,
# newness, company like and job like; taste is how close a role is to the
# ones the agent liked minus the ones it passed. cv_skills, cv_experience and
# cv_career {30, 25, 30} weight the judge's three numbers in the CV match.
# - limits, the analytics filter and the summarise size: {cut_fit_top: 500,
# cut_fit_min: 0.6, cut_cv_top: 150, cut_cv_min: 0.6, summarise_top: 30}, the
# count and the minimum by prompt closeness and by CV closeness, and how deep
# into each order summarise reads. See documentation/ADR Fetch prep and
# score.md.
# Two more, beside the starter in config.yaml:
agents: []
# - name: Remote from Europe
# prompts: