refactor
This commit is contained in:
Team3
@@ -164,27 +164,18 @@ EVIDENCE_CTX_LINES = 15 # context lines around a cited source position
|
||||
# ── Pipeline tuning (zentral, tunebar via CREATOR_PARAMS — siehe Override-Hook am Datei-Ende;
|
||||
# Registry mit Suchraum: backend/train_params.py). QA-/Detektor-Konstanten bleiben bewusst in
|
||||
# qa.py/guide_qa.py — die Messlatte darf nie Teil des Suchraums sein. ─────────────────────────
|
||||
SUBBLOCK_CHUNK = 10 # subblock finder: 1 agent per ~10 blocks, capped
|
||||
SUBBLOCK_MAX = 40 # chunk cap
|
||||
LEVEL_CHUNK = 100 # classifying is cheap → large packages
|
||||
RESEARCH_BATCH = 20 # crawl pages per batch
|
||||
RESEARCH_READERS = 2 # reader agents per batch/section (consensus ≥2)
|
||||
RESEARCH_THEMA_AGENTS = 5 # web mode (source "thema")
|
||||
RESEARCH_SECTION_CHARS = 12000 # uni/projekt section size (lost-in-the-middle guard)
|
||||
RESEARCH_RUNTIME = 900 # one research agent, one round (tail ingests live)
|
||||
CONSOLIDATION_CHUNK = 600 # up to here ONE global judge (fallback path)
|
||||
DEDUP_PAIR_FLOOR = 0.6 # min cosine for a candidate pair
|
||||
DEDUP_PAIRS_CHUNK = 40 # pairs per judge package
|
||||
DEDUP_TITLE_AUTO = 0.95 # near-identical TITLE cosine → merge without judge
|
||||
DEDUP_GLOBAL_FLOOR = 0.65 # global post-naming dedup candidate floor
|
||||
FILTER_CHUNK = 35 # blocks per judge in the degrade pass
|
||||
QUESTION_CHUNK_SUBS = 25 # target relevant subs per question chunk (LPT)
|
||||
QUESTION_MAX_ROUNDS = 3 # catch-up rounds for subs without a pattern
|
||||
FACTS_CHUNK_SUBS = 10 # facts extraction chunk (chunk count = parallelism)
|
||||
ARTEFACT_CHUNK_SUBS = 25 # flashcards/examples bulk chunk
|
||||
FACTS_CHECK_PANEL = 3 # judges per facts-check chunk (majority)
|
||||
CONSOLIDATION_PANEL = 3 # mapping judges per chunk
|
||||
SUBBLOCK_PANEL = 3 # judges in the subblock clarification
|
||||
# Board 2, verschmolzene Calls (block_calls.py): Panel-Größen der neuen Struktur.
|
||||
# Konsens braucht ≥2 unabhängige Nennungen bzw. Einstimmigkeit — 2 ist das Minimum,
|
||||
# 3 kauft Robustheit für +50 % Tokens auf dem jeweiligen Segment.
|
||||
@@ -193,6 +184,10 @@ VERIFY_PANEL = 2 # unabhängige Prüfer-Calls pro Block (+ Ersatz bei
|
||||
ART_SPLIT_SUBS = 20 # Artefakt-Generator splittet ab so vielen Subs in 2 parallele Calls
|
||||
FILTER_RECHECK_PANEL = 3 # judges in the survivor-recheck
|
||||
GATE_FIX_MIN = 3 # fact-gate: unbelegt-claims below min(this, relevante Subs) → log only (falsch fixt immer)
|
||||
ZIELE_MAX = 12 # Lernziele-Cap pro Block (mehr verwässert Coverage-Prüfung und Writer-Fokus)
|
||||
# Prüfer-Längenband um guide_qa.block_budget: enger als das QA-Band (0.35–1.5), damit der
|
||||
# Fix VOR der QA-Grenze greift. Die QA-Messlatte selbst bleibt bewusst in guide_qa.py.
|
||||
FIX_LAENGE_BAND = (0.5, 1.2)
|
||||
WRITER_SPLIT_SUBS = 30 # guide writer splits sections above this sub count
|
||||
KANBAN_BATCH = 5 # cards a worker pulls per micro-batch
|
||||
MAX_CARD_RETRIES = 3 # failures per card → dead-letter
|
||||
@@ -205,31 +200,22 @@ MAX_RESTARTS = 2 # agent restart cap per race slot
|
||||
# Fix-/Gate-Call (die laufen normal 110–135 s). 0 = aus.
|
||||
HEDGE_NACH_S = 90
|
||||
JUDGE_CHUNK = 40 # repair: findings per judge call
|
||||
EVENTS_RETENTION_TAGE = 60 # events älter als das werden beim Start gelöscht (Tabelle wuchs unbegrenzt)
|
||||
EVIDENCE_PER_BLOCK = 6000 # repair: excerpt chars per fremd candidate
|
||||
ABSCHLUSS_QA_LLM = 1 # 0 = Abschluss-QA ohne LLM-Judges (Training misst selbst; spart Minuten)
|
||||
|
||||
# Timeouts per agent step: (base seconds, seconds per block/section).
|
||||
# Applies equally to all providers — whoever is too slow gets restarted or overtaken.
|
||||
TIMEOUTS = {
|
||||
"research": (900, 0), # p95 measured 125 s (web mode); uni/link sections need headroom
|
||||
"research_mapping": (600, 3), # n = pre-merged entries
|
||||
"selection_mapping": (600, 2), # n = remaining entries (block inventory)
|
||||
"ergaenzung": (600, 0), # subject-field extension for projects (web research)
|
||||
"plan": (300, 5),
|
||||
"plan_judge": (600, 5), # judge reads up to 5 outlines, n = sections
|
||||
"content": (450, 30), # facts find/erg/fix — p95 measured 241 s (was 600+90n)
|
||||
# Judge caps tightened 2026-07-04: judge p50 is 6–72 s; a stalled call burns the whole
|
||||
# cap and its retry heals in seconds — the old 300 s base tripled the stall cost.
|
||||
"content_check": (150, 8), # content exam per block in the package
|
||||
"subblock": (400, 15), # finder round — p95 measured 124 s (was 900+45n)
|
||||
"subblock_check": (150, 10), # judge decides contested subblocks in the chunk
|
||||
"konsolidierung": (300, 20), # consolidation judge sees ALL subs with key points
|
||||
"level": (300, 10), # classify subblocks per chunk
|
||||
"level_check": (150, 8), # judge decides contested levels in the chunk
|
||||
"relevance": (300, 10), # subblocks relevant/peripheral per chunk
|
||||
"relevance_check": (150, 8), # judge decides contested relevance in the chunk
|
||||
"question_pattern": (300, 15), # question patterns per block (subblocks × types)
|
||||
"question_pattern_check": (150, 8), # critic cleans up the pattern table per block
|
||||
# Board 2, verschmolzene Calls: größere Outputs pro Call, dafür wenige Segmente
|
||||
"generate": (450, 0), # Subs+Facts+Level in einem (Sub-Zahl vorab unbekannt)
|
||||
"verify": (300, 10), # Audit über alle Subs (n = Subs), key points gekappt
|
||||
@@ -237,19 +223,11 @@ TIMEOUTS = {
|
||||
"artefakt": (450, 15), # Fragen+Karten+Beispiele (n = Subs)
|
||||
"artefakt_check": (200, 8), # Beispiel-Verifikation + Fragen-Kritik (n = Subs)
|
||||
"writer": (450, 60), # per section — split keeps sections ≤30 subs
|
||||
"lese_check": (300, 10), # per section in the package
|
||||
# guide board (per card = one block)
|
||||
"lernziele": (300, 5), # backward-design objectives per block
|
||||
"fakten_gate": (600, 5), # CoVe claim check per block
|
||||
"coverage": (300, 5), # objective↔section mapping per block
|
||||
}
|
||||
|
||||
# Purpose per format — flows into the outline judge (what the guide should achieve).
|
||||
# German strings: these are inserted verbatim into the judge prompt → kept German on purpose.
|
||||
FORMAT_PURPOSE = {
|
||||
"Guide": "einen fokussierten Guide — alles Relevante ohne Randthemen",
|
||||
"FullGuide": "einen Komplett-Guide — das ganze Thema inkl. Randthemen",
|
||||
"Rest": "einen Ergänzungs-Guide — nur die Randthemen",
|
||||
"pruefer": (600, 5), # verschmolzener Qualitäts-Pass (Gate+Coverage+Lese) per block
|
||||
# QA/Repair-Judge-Wellen (qa.judge_wave) — außerhalb der Boards, keine n-Skalierung
|
||||
"qa_judge": (600, 0),
|
||||
}
|
||||
|
||||
# Provider stacks: completely independent, any one can be removed at any time.
|
||||
|
||||
Reference in New Issue
Block a user