Configuration

Use config.example.yaml as a base and pass your file with --config.

bm25_min_score: 0.30
overlap_min_jaccard: 0.15
unsupported_ratio_threshold: 0.25
top_k: 5
top_k_max: 10
context_reduction_mode: chunk

# Generation providers
anthropic_model: claude-sonnet-4-20250514
openai_model: gpt-4o-mini
max_tokens: 1024
stream_generation: false

# Hybrid retrieval
embedding_model: sentence-transformers/all-MiniLM-L6-v2
rrf_k: 60

# Production EGA verifier
nli_model: cross-encoder/nli-deberta-v3-small
nli_batch_size: 8

Key	Type	Default	Description
`bm25_min_score`	`float`	`0.30`	Minimum BM25 score required for confident retrieval.
`overlap_min_jaccard`	`float`	`0.15`	Minimum Jaccard overlap for stub EGA sentence support.
`unsupported_ratio_threshold`	`float`	`0.25`	Maximum unsupported sentence ratio before grounding fails.
`top_k`	`int`	`5`	Number of retrieved chunks requested per query.
`top_k_max`	`int`	`10`	Upper bound for retrieval expansion repair.
`context_reduction_mode`	`str`	`chunk`	Context reduction strategy during grounding repair.
`anthropic_model`	`str`	`claude-sonnet-4-20250514`	Anthropic model id for generation provider.
`openai_model`	`str`	`gpt-4o-mini`	OpenAI model id for generation provider.
`max_tokens`	`int`	`1024`	Maximum generated tokens for API providers.
`stream_generation`	`bool`	`false`	Enables streaming generation for API providers.
`embedding_model`	`str`	`sentence-transformers/all-MiniLM-L6-v2`	Dense embedding model for hybrid retrieval.
`rrf_k`	`int`	`60`	Reciprocal Rank Fusion constant for BM25 + dense ranks.
`nli_model`	`str`	`cross-encoder/nli-deberta-v3-small`	Cross-encoder NLI model for production EGA verifier.
`nli_batch_size`	`int`	`8`	Batch size for sentence/chunk NLI scoring.