diff --git a/docs/ranking-flow.md b/docs/ranking-flow.md index 3bdfa9a..640bd94 100644 --- a/docs/ranking-flow.md +++ b/docs/ranking-flow.md @@ -2,6 +2,8 @@ This is how a role gets from a board to the Roles feed, for one search agent. It is the shared picture to change the flow against. Each stage says what the code does **today**, what was **decided** on 2026-10-04 (not built yet) and what is still **open**. +Names: **fetch text** is the second HTTP read of a role's own page, with no model (the code calls it `enrich`). **Extract**, **judge** and **summarise** are the model calls. + ``` company boards (ATS) search providers reach: Where lines + fetch terms, their own saved query @@ -14,10 +16,10 @@ This is how a role gets from a board to the Roles feed, for one search agent. It 2 Cheap pass, no tokens ─ prep, embed, fit to the prompts, closeness to the CV, gates │ ▼ analysis score → the cut (at most N, each at least X) - 3 Second fetch ─ enrich reads the role's own page for cut roles without text → embed, rescore + 3 Fetch text ─ the role's own page, for cut roles without text → embed, rescore │ ▼ - 4 Model pass (Haiku) ─ extract, summarise, judge, on the cut only + 4 Model pass (Haiku), the whole cut ─ extract → rescore → judge, then summarise │ ▼ 5 Feed ─ two orders: "what I'm looking for" (final score) and "what fits me" (CV match) @@ -60,11 +62,11 @@ Without the Where lines, about four in five model reads go to roles outside DK, - How affinity enters: as a weighted term, or by widening the cut for liked companies. - Whether a role liked or saved outside the cut is pushed into the model pass. Today saved and liked roles are summarised whatever their score. -## 3 Second fetch +## 3 Fetch text -**Today.** Enrich reads the role's own page for open roles without text, on hosts that list without it. It uses its own bar from the config (`enrich.top` 500, `enrich.min_score` 0.5, at most 200 per board and 1,000 per run). Embed and score run again for the roles it filled. +**Today.** The `enrich` stage reads the role's own page for open roles without text, on hosts that list without it. It uses its own bar from the config (`enrich.top` 500, `enrich.min_score` 0.5, at most 200 per board and 1,000 per run). Embed and score run again for the roles it filled. -**Decided.** Enrich reads the analysis cut, so every role the model will read has its text first. +**Decided.** Fetch text reads the whole analysis cut, so every role the model will read has its text first. The stage is renamed from `enrich` to fetch text so it is not mistaken for a model call. ## 4 Model pass @@ -74,7 +76,16 @@ Without the Where lines, about four in five model reads go to roles outside DK, - **Summarise** reads at most 30 roles scoring at least 0.65, plus the roles you saved or liked. - **Judge** reads at most 150 roles scoring at least 0.55, plus the 50 closest to the CV that those leave out. It scores skills, experience and career 0 to 100, plus a verdict on each must-have. These are weighted 30/25/30 into `cv_score`. -**Decided.** The form wording becomes "at most N, each scoring at least X", in place of "the best N down to a score of X". +**Decided.** One depth: the analysis cut is the only gate, and every model stage reads all of it. A second score inside the cut would defeat the purpose of the cut. The order is: + +1. **Extract** reads the whole cut. It is the small, fast call (about 1.5 s a role on the live backend). +2. A rescore applies the gates to what extract filled, mainly location and remote area. The cut is the same gate as before, recomputed with better facts. A role that a rescore moves into the cut is extracted on the next pull. +3. **Judge** reads the cut as recomputed (about 6 s a role). +4. **Summarise** is an afterburner. It does not feed any score. + +The form wording for the sizes becomes "at most N, each scoring at least X", in place of "the best N down to a score of X". + +**Cost.** Every model read happens once per role (judge once per role and CV), so a larger cut costs mostly a one-off catch-up. With a cut of about 650 roles, that catch-up is about 16 minutes of extract and about an hour of judge on the live backend (`claude_code`, the subscription, so no per-token bill). After it, each pull reads only the roles new to the cut. **Open.** Whether extract and judge become one call. Arguments for keeping two: extract is per role and needs no CV, while judge is per agent and role with the CV in the system prompt; and extract's location and remote answers feed the gates before judge reads anything. The argument for one call is that the posting is read once instead of twice.