xrag-temporal

Time-aware uncertainty for RAG — temporal decay of evidence confidence using Subjective Logic.

Your RAG system treats every retrieved document as equally valid — whether it was written yesterday or three years ago. xrag-temporal fixes this by applying subjective logic temporal decay: as evidence ages, belief migrates into uncertainty.

Install

pip install xrag-temporal

Quick Start

from xrag_temporal import opinion, decay, decay_series, should_abstain

# Create an opinion from evidence
ev = opinion(belief=0.8, disbelief=0.1)  # strong evidence

# Decay it: 30 days old, 7-day half-life
stale = decay(ev, elapsed_days=30, half_life_days=7)
print(stale)
# → Opinion(b=0.0442, d=0.0055, u=0.9503)  — almost all uncertainty!

# Should the system abstain?
print(should_abstain(stale))  # True — uncertainty > 0.7

The Problem

Standard RAG pipelines retrieve documents and feed them to an LLM without considering when the evidence was created. A 3-year-old article about "the CEO of Twitter" will confidently produce the wrong answer.

The Solution

xrag-temporal decays belief and disbelief into uncertainty as evidence ages:

• Fresh evidence (age ≈ 0): original confidence preserved
• Aging evidence: belief/disbelief shrink, uncertainty grows
• Stale evidence: nearly all mass is uncertainty → system knows it doesn't know

Three decay functions:

• Exponential (default): smooth, never fully zero — λ = 2^(-t/τ)
• Linear: reaches zero at 2× half-life — λ = max(0, 1 - t/2τ)
• Step: binary fresh/stale — λ = 1 if t < τ else 0

License

MIT