WuJunde commited on
Commit
95e69fa
·
1 Parent(s): 79b11fa

new version: med-postrag

Browse files
Files changed (2) hide show
  1. Dockerfile +1 -1
  2. README.md +46 -6
Dockerfile CHANGED
@@ -1 +1 @@
1
- FROM jundewu/medgraphrag-pubmedonline:latest
 
1
+ FROM jundewu/medrag-post
README.md CHANGED
@@ -1,10 +1,50 @@
1
  ---
2
- title: Medrag
3
- emoji: 🔥
4
- colorFrom: pink
5
- colorTo: blue
6
  sdk: docker
7
- pinned: false
8
  ---
9
 
10
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ title: MedGraphRAG PubMed
3
+ emoji: 🏥
4
+ colorFrom: blue
5
+ colorTo: green
6
  sdk: docker
7
+ app_port: 7860
8
  ---
9
 
10
+ # MedRAG PubMed Online
11
+
12
+ This is a RAG-based medical research assistant that searches PubMed literature to answer the medical questions. Currently doing RAG based on literature titles.
13
+
14
+ ## Run locally
15
+ pip install -r requirements.txt (see Dockerfile for more details)
16
+
17
+ export OPENAI_API_KEY= "your-openai-api-key"
18
+ export NCBI_API_KEY= "your-ncbi-api-key" (optional)
19
+
20
+ python rerankrag.py/simplerag.py "your-medical-research-question"
21
+
22
+ ## Todo list:
23
+
24
+ - [x] integrate pubmed search progress into the citation pipeline
25
+ - [x] rerank and cutting-edge langchain retriever
26
+ - [x] gradio interface
27
+ - [ ] include the abstract/full text in retrieval (efficiency?)
28
+ - [ ] sophasticated search key word of pubmed
29
+ - [ ] refine answer based on searched literatures
30
+ - [ ] make it a non-accessable container and release to huggingface
31
+ - [ ] consider the weight of literatures (IF?)
32
+ - [ ] use Medline (med wiki website) to provide grounded explanation and definition
33
+ - [ ] discriminate different kind of med questions (diagnosis, learn med knowledge, drug match…) and process with different agent-chains
34
+ - [ ] hard code literature index
35
+
36
+ ## Customer feedback (from health-eco, 2024-12-07)
37
+ - [x] The answers are too generic - need to extract and show specific treatment plans, disease mechanisms, and patient types from the medical papers we find (updated a version a solve this)
38
+ - [ ] Need to be honest about limitations - say "I don't know" clearly when we can't find good answers in our search results
39
+ - [ ] first check a local knowledge graph we maintain, then fall back to online PubMed searches if needed. Keep updating the local graph with new findings
40
+ - [ ] Add a way to rate how trustworthy each source is, maybe looking at things like the journal's Impact Factor, how often it's cited, the journal's reputation, when it was published, and who wrote it
41
+
42
+ ## Customer feedback (from us-insurance, 2024-12-09)
43
+ - [ ] Recommend clinical services to patients based on their queries and identify the most cost-efficient, nearest clinic within their insurance network using the data available at: https://transparency-in-coverage.uhc.com.
44
+
45
+ ## Customer feedback (from AI-EHR, 2024-12-09)
46
+ - [ ] cheaper than GPT
47
+ - [ ] no latency (they provide phone service)
48
+ - [ ] diagnosis based on the hospital's private guidelines (personized)
49
+ - [ ] better searching capabilties based on patient's medical history
50
+ - [ ] screen monitoring (mainly OCR) based co-pilot (avoid regulation concern as user don't need to upload their data)