82 lines
2.8 KiB
YAML
82 lines
2.8 KiB
YAML
|
# Example routing config for a trading-literature Lefevre-style run.
|
||
|
|
#
|
||
|
|
# Captures the IB-WP-0018 task-type taxonomy from docs/routing-task-types.md:
|
||
|
|
# summarize-source → cheap model (volume-heavy, recoverable downstream)
|
||
|
|
# extract-entities → smart model (durable output; be strict)
|
||
|
|
# extract-relations → smart model (depends on entities)
|
||
|
|
# evaluate-entity → judge model (different family from extraction)
|
||
|
|
# synthesize-report → smart model (volume-of-one, quality matters, cheap)
|
||
|
|
#
|
||
|
|
# Quality floors are the recommended starting points from
|
||
|
|
# docs/routing-task-types.md. With a ledger configured, AdaptiveRoutingPolicy
|
||
|
|
# will pick the cheapest *qualifying* adapter per task type as observations
|
||
|
|
# accumulate; until then it falls back to the static prefer/fallback order.
|
||
|
|
#
|
||
|
|
# Refresh the model rates in src/infospace_bench/model_rates.yaml before any
|
||
|
|
# full-book run — list prices drift, and the rough USD estimate in the budget
|
||
|
|
# log depends on them.
|
||
|
|
|
||
|
|
schema_version: 1
|
||
|
|
|
||
|
|
# Workspace-relative ledger so QualityLedger observations from this workspace
|
||
|
|
# stay with this workspace. Drop this line to run pure static routing.
|
||
|
|
ledger_path: output/routing/quality.jsonl
|
||
|
|
|
||
|
|
# Floors apply when --quality-floor is not passed at the call site. The CLI
|
||
|
|
# flag wins, then the per-task quality_floor below, then this default.
|
||
|
|
default_quality_floor: 0.80
|
||
|
|
|
||
|
|
stage_to_task_type:
|
||
|
|
summarize-source: cheap
|
||
|
|
extract-entities: smart
|
||
|
|
extract-relations: smart
|
||
|
|
evaluate-entity: judge
|
||
|
|
synthesize-report: smart
|
||
|
|
|
||
|
|
task_types:
|
||
|
|
|
||
|
|
cheap:
|
||
|
|
quality_floor: 0.70
|
||
|
|
candidates:
|
||
|
|
- id: openrouter:gpt-4o-mini
|
||
|
|
provider: openrouter
|
||
|
|
model: openai/gpt-4o-mini
|
||
|
|
api_key_env: OPENROUTER_API_KEY
|
||
|
|
max_cost_per_1k: 0.001
|
||
|
|
- id: openrouter:claude-3.5-haiku
|
||
|
|
provider: openrouter
|
||
|
|
model: anthropic/claude-3.5-haiku
|
||
|
|
api_key_env: OPENROUTER_API_KEY
|
||
|
|
max_cost_per_1k: 0.003
|
||
|
|
|
||
|
|
smart:
|
||
|
|
quality_floor: 0.85
|
||
|
|
candidates:
|
||
|
|
- id: openrouter:claude-3.5-haiku
|
||
|
|
provider: openrouter
|
||
|
|
model: anthropic/claude-3.5-haiku
|
||
|
|
api_key_env: OPENROUTER_API_KEY
|
||
|
|
- id: openrouter:claude-3.5-sonnet
|
||
|
|
provider: openrouter
|
||
|
|
model: anthropic/claude-3.5-sonnet
|
||
|
|
api_key_env: OPENROUTER_API_KEY
|
||
|
|
|
||
|
|
judge:
|
||
|
|
quality_floor: 0.80
|
||
|
|
candidates:
|
||
|
|
# Evaluation goes through a different family than extraction to limit
|
||
|
|
# self-preference bias.
|
||
|
|
- id: openrouter:gpt-4o-mini
|
||
|
|
provider: openrouter
|
||
|
|
model: openai/gpt-4o-mini
|
||
|
|
api_key_env: OPENROUTER_API_KEY
|
||
|
|
|
||
|
|
# Baseline is wired here so a follow-up T05 ShadowingAdapter step can
|
||
|
|
# reference `claude-code` as the grading oracle without editing the
|
||
|
|
# task_types stanza.
|
||
|
|
baseline:
|
||
|
|
candidates:
|
||
|
|
- id: claude-code
|
||
|
|
provider: claude_code
|
||
|
|
model: claude-opus-4-7
|