Reference

Adaptive retrieval and compression¶

Behavior¶

Mechanism	Mari decides	Evidence boundary
CRAG	Use, supplement, or replace retrieved context from score thresholds	SciFact gold relevance exercised all three routes. The relevance evaluator remains yours
FLARE	Retrieval trigger for low-confidence future tokens	QASPER answer novelty exercises masking. Real calibration comes from your model
Self-RAG	Candidate utility and whether retrieval is requested	Gold relevance validates score composition. Reflection-token prediction needs a separate measure
Chain-of-Note	Retrieved, parametric, or unknown answer source	Gold qrels validate deterministic source choice. Note judgment quality needs a separate measure
RECOMP	Which complete sentences fit a token budget	QASPER retained `16.6%` of tokens with evidence recall `0.330` in the measured configuration

Example routing examples

Signal pattern	Decision	Next operation
High relevance	Accept	Use retrieved corpus evidence
Ambiguous relevance	Supplement	Add an external retrieval arm
Low relevance	Replace	Discard current candidates and retrieve again
Low-confidence future tokens	Retrieve	Mask uncertain tokens and search with the remainder

Retrieval can be triggered, revised, rescored, or compressed at explicit decision points across a model call.

How it works¶

CRAG routing maps evaluator scores through two thresholds to use the corpus, augment it with external search, or replace it. FLARE finds low-confidence tokens in a predicted future sentence, removes those tokens, and uses the remaining text as a retrieval query before regeneration. Self-RAG combines generation, retrieval, relevance, support, and utility signals with visible weights. RECOMP selects scored sentences under a token budget, then restores their source order.

Papers

Self-RAG: reflection-token scoring CRAG: corrective retrieval FLARE: forward-looking active retrieval RECOMP: selective compression

adaptive_retrieval.py · current¶

from mari_kit.retrieval import (
    CompressionSentence, plan_active_retrieval, plan_corrective_retrieval,
    selective_compression,
)
from mari_kit.verification import score_self_rag_candidate

correction = plan_corrective_retrieval(retrieval_evaluator(query, hits),
    lower_threshold=-0.8, upper_threshold=0.6)
active = plan_active_retrieval(future.tokens, future.probabilities, threshold=0.2)
reflection = score_self_rag_candidate(generation_probability=signals.generation,
    retrieve_probability=signals.retrieve, relevance_probability=signals.relevant,
    support_probability=signals.supported, utility=signals.utility)

compressed = selective_compression([
    CompressionSentence(sentence_id=s.id, text=s.text, token_count=count_tokens(s.text),
        relevance=compressor.score(query, s.text)) for s in sentences
], token_budget=600, relevance_threshold=0.4)