Bitwise Knowledge Assistant

Why Document Ranking Matters

Not all documents are equally reliable or current. A prospectus from 2024 should take precedence over a marketing draft from 2022. Authority and recency scoring helps the system prioritize the most trustworthy and up-to-date information when answering your questions.

Authority Scoring

Authority scoring assigns different weights to documents based on their type and source. Official documents like prospectuses and fact sheets receive higher scores than informal documents like meeting notes or drafts.

Document Authority Tiers

Document Type	Authority	Weight
Prospectus, SAI	Highest	1.0
Fact Sheet, Product Doc	High	0.9
Research Report	Medium-High	0.8
Investment Memo	Medium	0.7
Presentation	Medium	0.6
Internal Memo	Lower	0.5
Draft, Notes	Low	0.3

How Authority is Determined

Authority is assigned during document ingestion based on:

File path: Documents in certain folders receive higher authority
File name: Keywords like “prospectus” or “fact_sheet” boost authority
Classification: The AI document classifier contributes to authority
Metadata: Explicit authority tags in document metadata

Recency Scoring

Recency scoring boosts newer documents over older ones. This is particularly important for time-sensitive information like market data, performance figures, and regulatory updates.

Recency Decay Model

Document scores decay over time using an exponential model. The decay rate is configurable but defaults to a half-life of approximately 180 days.

Today: 1.0 (full score)

6 months ago: ~0.5 (half score)

1 year ago: ~0.25 (quarter score)

2+ years ago: ~0.1 (minimum floor)

When Recency Matters Most

Performance data: Returns, prices, market values
AUM and flows: Fund sizes change frequently
Regulatory status: SEC approvals, compliance updates
Market commentary: Analysis tied to current conditions

When Recency Matters Less

Product structure: ETF mechanics don't change often
Educational content: Fundamental explanations stay valid
Historical analysis: Past events remain accurate

How Scores Combine

The final relevance score for a document combines three factors:

Combined Score Formula

final_score = semantic_score × authority_weight × recency_weight

For example, a highly relevant (0.8) prospectus (1.0 authority) from last month (0.95 recency) would score 0.8 × 1.0 × 0.95 = 0.76

Example Comparison

Document	Semantic	Authority	Recency	Final
2024 Prospectus	0.75	1.0	0.9	0.68
2022 Research Report	0.85	0.8	0.3	0.20
Recent Draft Memo	0.90	0.3	0.95	0.26

In this example, the 2024 Prospectus wins despite having lower semantic similarity because its high authority and recent date outweigh the research report's slightly better semantic match.

Authority Configuration

Authority weights are configured in the assets/authority_config.json file. Administrators can adjust weights for different document types, paths, and keywords.

Configuration Options

path_patterns: Boost documents in specific folders
filename_patterns: Boost by filename keywords
doc_type_weights: Set weights by classification type
recency_half_life: Control how fast recency decays
min_recency_score: Floor for very old documents

Impact on Search Results

Authority and recency scoring affects which documents appear in your results and in what order. When multiple documents discuss the same topic:

Official documents rank higher than drafts
Recent documents rank higher than old ones
A very relevant old draft may still rank below a moderately relevant new prospectus

Retrieval Settings Guide - Configure relevance thresholds
Document Ingestion Explained - How documents are classified