Spaces:
Sleeping
Sleeping
EMMA commited on
Commit ·
e159cd3
1
Parent(s): 2bf1da0
Add tiktoken cache to Dockerfile and setup GitHub Actions CI/CD
Browse files
README.md
CHANGED
|
@@ -1,3 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# EMMA — Empathetic Memory-Augmented Multi-layer Assistant
|
| 2 |
*(Research Prototype)*
|
| 3 |
|
|
@@ -11,12 +22,20 @@ This repository provides a reference implementation of a mobile-friendly, memory
|
|
| 11 |
---
|
| 12 |
|
| 13 |
## Table of Contents
|
| 14 |
-
- [
|
| 15 |
-
- [
|
| 16 |
-
- [
|
| 17 |
-
- [
|
| 18 |
-
- [
|
| 19 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
---
|
| 22 |
|
|
@@ -40,14 +59,10 @@ This repository provides a reference implementation of a mobile-friendly, memory
|
|
| 40 |
- **Evaluation tooling**:
|
| 41 |
Scripts supporting quantitative memory retrieval accuracy and qualitative Likert-scale evaluation pipelines.
|
| 42 |
|
| 43 |
-
|
| 44 |
-
|
| 45 |
---
|
| 46 |
|
| 47 |
## Architecture (High Level)
|
| 48 |
|
| 49 |
-
|
| 50 |
-
|
| 51 |
- **Indexing**: Episodic and semantic memory items are embedded and stored in vector indexes.
|
| 52 |
- **Routing**: A classifier determines which memory layer(s) should be queried. Hybrid queries may combine episodic and semantic retrieval.
|
| 53 |
- **Prompting**: Retrieved memory is merged into therapy-aware prompt templates prior to LLM invocation.
|
|
@@ -118,9 +133,3 @@ Evaluation can be reproduced by preparing:
|
|
| 118 |
- Automated evaluators and LLM judgments may be noisy or biased; safety-critical use cases require clinician oversight and human-in-the-loop validation.
|
| 119 |
- The system may occasionally hallucinate memory-grounded facts; retrieval traces should always be logged for auditing and debugging.
|
| 120 |
- See the associated paper for a detailed discussion of limitations and evaluation methodology.
|
| 121 |
-
|
| 122 |
-
---
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
|
|
|
| 1 |
+
---
|
| 2 |
+
title: Emma
|
| 3 |
+
emoji: 🧠
|
| 4 |
+
colorFrom: blue
|
| 5 |
+
colorTo: indigo
|
| 6 |
+
sdk: gradio
|
| 7 |
+
sdk_version: 4.44.0
|
| 8 |
+
app_file: app.py
|
| 9 |
+
pinned: false
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
# EMMA — Empathetic Memory-Augmented Multi-layer Assistant
|
| 13 |
*(Research Prototype)*
|
| 14 |
|
|
|
|
| 22 |
---
|
| 23 |
|
| 24 |
## Table of Contents
|
| 25 |
+
- [EMMA — Empathetic Memory-Augmented Multi-layer Assistant](#emma--empathetic-memory-augmented-multi-layer-assistant)
|
| 26 |
+
- [Table of Contents](#table-of-contents)
|
| 27 |
+
- [Key Features](#key-features)
|
| 28 |
+
- [Architecture (High Level)](#architecture-high-level)
|
| 29 |
+
- [Personalized Response Generation Workflow](#personalized-response-generation-workflow)
|
| 30 |
+
- [Step 1 – User Query](#step-1--user-query)
|
| 31 |
+
- [Step 2 – Query Classification](#step-2--query-classification)
|
| 32 |
+
- [Step 3 – Memory Routing](#step-3--memory-routing)
|
| 33 |
+
- [Step 4 – Memory Retrieval](#step-4--memory-retrieval)
|
| 34 |
+
- [Step 5 – Prompt Composition](#step-5--prompt-composition)
|
| 35 |
+
- [Step 6 – Response Generation](#step-6--response-generation)
|
| 36 |
+
- [Evaluation \& Metrics](#evaluation--metrics)
|
| 37 |
+
- [Reproducibility](#reproducibility)
|
| 38 |
+
- [Limitations \& Safety](#limitations--safety)
|
| 39 |
|
| 40 |
---
|
| 41 |
|
|
|
|
| 59 |
- **Evaluation tooling**:
|
| 60 |
Scripts supporting quantitative memory retrieval accuracy and qualitative Likert-scale evaluation pipelines.
|
| 61 |
|
|
|
|
|
|
|
| 62 |
---
|
| 63 |
|
| 64 |
## Architecture (High Level)
|
| 65 |
|
|
|
|
|
|
|
| 66 |
- **Indexing**: Episodic and semantic memory items are embedded and stored in vector indexes.
|
| 67 |
- **Routing**: A classifier determines which memory layer(s) should be queried. Hybrid queries may combine episodic and semantic retrieval.
|
| 68 |
- **Prompting**: Retrieved memory is merged into therapy-aware prompt templates prior to LLM invocation.
|
|
|
|
| 133 |
- Automated evaluators and LLM judgments may be noisy or biased; safety-critical use cases require clinician oversight and human-in-the-loop validation.
|
| 134 |
- The system may occasionally hallucinate memory-grounded facts; retrieval traces should always be logged for auditing and debugging.
|
| 135 |
- See the associated paper for a detailed discussion of limitations and evaluation methodology.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|