How Aura Works

Aura is a clinical mental health assistant powered by a hybrid search pipeline, cross-encoder reranking, and structured clinical data from the DSM-5-TR. Every architectural decision is measured against a 150-query evaluation harness.

Search Pipeline

Each query passes through query expansion, dual-path retrieval (semantic + keyword), cross-encoder reranking, and LLM generation with constitutional safety principles.

Query

User input

0ms

Expansion

80+ synonyms

~1ms

Embedding

MiniLM-L6-v2

~50ms

Hybrid Search

Vector + BM25

~20ms

Reranking

Cross-encoder

~300ms

LLM

Llama 3.3 70B

~1.5s

Safety

Constitutional

In-prompt

Response

SSE stream

~2s total

Retrieval Benchmarks

Measured against 150 gold Q&A pairs across 5 categories: scope, clinical depth, differential diagnosis, safety, and edge cases.

Run the eval harness to generate benchmark data:

npm run eval

System Architecture

Knowledge Base

594

Articles

8,753

Search chunks

DSM-5-TR disorders

Personality disorders

Search Pipeline

384d

Embedding dimensions

60/40

Vector / BM25 weight

ms-marco-MiniLM-L-6-v2

Reranker model

20→8

Candidates → results

Safety Pipeline

3-tier

Safety classification

988

Crisis escalation

20/min

Rate limit burst

In-prompt

Constitutional principles

Model Stack

Llama 3.3 70B

Generation model

all-MiniLM-L6-v2

Embedding model

Groq

Inference provider

0.3

Temperature

Method Comparison

Retrieval quality across search methods, measured on queries with known expected sources.

Method	Recall@3	Recall@5	Recall@8	MRR	NDCG@10
BM25 Only	—	—	—	—	—
Hybrid (Vector + BM25)	89.9%	92.0%	93.1%	87.7%	87.9%
Hybrid + Reranking	—	—	—	—	—

Knowledge Graph

Interactive visualization of relationships between conditions, screening tools, and diagnostic categories extracted from structured clinical data.

Loading knowledge graph...

Contact

Interested in our research or data? Reach out at contact@moodspan.org.