Module rpe 

D-MEM RPE-based tiered graph extraction routing.

Computes a heuristic “reward prediction error” (RPE) signal for each incoming turn. Low-RPE turns (predictable, topically continuous, no new entities) skip the expensive MAGMA LLM extraction pipeline. High-RPE turns proceed to full extraction.

RPE formula

RPE = 0.5 * (1 - max_cosine_similarity) + 0.5 * entity_novelty_ratio

Where:

• max_cosine_similarity = max cosine similarity between current turn embedding and last N turn embeddings. High = topically predictable.
• entity_novelty_ratio = fraction of candidate entity names not seen in recent history. 0.0 if no entities extracted.

Safety valve

To prevent unbounded skipping, consecutive_skips is tracked. When it reaches max_skip_turns, extraction is forced regardless of RPE score.

Structs

RpeRouter

Stateful RPE router. Tracks recent embeddings and entity history.

RpeSignal

RPE computation result for a single turn.

Constants

RPE_EMBEDDING_BUFFER_SIZE

Maximum number of recent turn embeddings to keep for context similarity computation.

Functions

extract_candidate_entities

Extract candidate entity names from text using simple heuristics.