5.4.15.42 combined chatgpt-turbo-16k and search term prompt identification feature. additionally aligned with the cloud data structure for seperation of fields and prompts, added ai formating instruction and simlified ui.

app.py

@@ -18,9 +18,10 @@ examples = []
 # app_theme = gr.themes.Base(neutral_hue="blue")
 
 demo = gr.TabbedInterface(
-    [device_page,study_page,equivalent_page], 
-    ["Device","Clinical Study Report","Equivalent Comparators"],
-    # theme= app_theme,
+    # [device_page,study_page,equivalent_page], 
+    # ["Device","Clinical Study Report","Equivalent Comparators"],
+    [study_page],
+    ["Clinical Study"],
     theme = gr.themes.Soft(primary_hue="sky",secondary_hue="orange"),
     css = "footer {visibility: hidden}",
     title="AMRA AI Medi Reader")

application.py

@@ -57,7 +57,8 @@ data_structure = {
         "term",
         "clinical term",
         "summary term",
-        "template_name"
+        "template_name",
+        "terms"
     ]},
     "prompts":{
         "key":[
@@ -72,6 +73,8 @@ data_structure = {
             "levels",
             "preoperatives",
             "prompt",
+            "fields",
+            "reformat_inst"
         ]
     },
     "articles":{
@@ -92,6 +95,7 @@ data_structure = {
         "key":[
             "domain",
             "article",
+            "outcomes"
         ],
         "fields":[
             "domain",
@@ -105,12 +109,8 @@ data_structure = {
 application default data
 '''
 app_data = {
-    "articles":[],
+    "current_article":{},
     "terms":[],
     "prompts":[],
     "outputs":[]
 }
-
-
-# hypothesis:
-# normal abstract length is about 800 characters

cloud_storage.py

@@ -48,11 +48,11 @@ def upload_fileobj(file_obj, bucket, object_name=None):
         object_name = file_obj.name
 
     try:
-        s3.upload_fileobj(file_obj, bucket, object_name)
+        res = s3.upload_fileobj(file_obj, bucket, object_name)
     except Exception as e:
         print(e)
-        return False
-    return True
+        return e
+    return res
 
 # get a file from s3
 def download_file(bucket, object_name, file_name=None):

features.py

@@ -1,8 +1,5 @@
 # language default packages
 from datetime import datetime
-from operator import mul
-from functools import reduce
-from sys import stdout
 from collections import defaultdict
 
 # external packages
@@ -29,10 +26,6 @@ def init_app_data():
 def process_study(
         study_file_obj,
         study_content,
-        performance_metric_1, 
-        performance_metric_2, 
-        safety_metric_1, 
-        safety_metric_2, 
         device=default_device
         ):
     
@@ -43,89 +36,69 @@ def process_study(
     else:
         return "No file or content provided","No file or content provided","No file or content provided"
 
-    prompts = select_prompts( # need to identify how the app will know which prompts to use
-        article,
-        performance_metric_1,
-        performance_metric_2,
-        safety_metric_1,
-        safety_metric_2
-        
-    )
-    # print("check prompts",prompts)
+    app_data["current_article"] = article
+    selected_prompts = select_prompts(article["content"],terms=app_data["terms"],prompts=app_data["prompts"])
     
     output = {
         "domain":article["domain"],
         "article":article["name"],
-        "output":defaultdict(dict)
+        "outcomes":defaultdict(str)
     }
 
-    for p in prompts:
-        prompt_string = ""
-        for s in p["sections"].split(","):
-            prompt_string += f"{article[s]}"
-
-        prompt_string += f"\n {p['prompt']}"
-        with open(f".prompts/{article['name']}_{p['template_name']}.txt","w") as f:
-            f.write(prompt_string)
-
-        res = execute_prompt(prompt_string)
-
-        with open(f".outputs/{article['name']}_{p['template_name']}.txt","w") as f:
-            f.write(res)
-
-        output["output"][p["assessment_step"]][p["template_name"]]=res
-
-    
+    res = process_prompts(article["content"],selected_prompts)
+    output["outcomes"] = res
         
-    overview = create_overview(output["output"]["overview"])
-    details = create_details(output["output"])
+    # overview = create_overview(output["output"]["Overview"])
+    views = create_views(res)
 
     add_output(output)
 
-    return overview, details
+    return views
+    # return ""
 
 def refresh():
     '''
     this function refresh the application data from the cloud backend
     '''
     init_app_data()
-    return "refreshed", "refreshed"
 
-def create_overview(overview_list):
-    '''
-    '''
-    md_text = "## Overview\n\n"
-    md_text += "| attributes | detail |\n|:---|:---|\n"
-    for _,v in overview_list.items():
-        r = v.replace("\n\n","")
-        rows = r.split("\n")
-        for r in rows:
-            c = r.replace(": "," | ")
-            md_text += f"| {c} |\n"
-    # with open("overview.md","w") as f:
-        # f.write(md_text)
-    return gr.update(value=md_text)
+    article = app_data["current_article"]
+    selected_prompts = select_prompts(article["content"],terms=app_data["terms"],prompts=app_data["prompts"])
+    
+    output = {
+        "domain":article["domain"],
+        "article":article["name"],
+        "outcomes":defaultdict(str)
+    }
 
-def create_details(output):
-    sections = ["clinical", "radiographic", "fussion assessment", "other","safety"]
-    titles = ["Clinical Outcomes", "Radiological Outcomes", "Fussion Assessment", "Other Outcomes","Safety Outcomes"]
+    res = process_prompts(article["content"],selected_prompts)
+    output["outcomes"] = res
+    views = create_views(res)
+    add_output(output)
+
+    return views
 
+def create_views(output):
     md_text = ""
-    for section, title in zip(sections, titles):
-        md_text += f"## {title}\n\n"
-        
-        for key,table in output[section].items():
-            md_text += f"### {key} \n\n"
-            rows = table.replace("\n\n","").split("\n")
-            for i,r in enumerate(rows):
-                cells = r.split("\t")
-                md_text += f"| {' | '.join(cells)} |\n"
-                if i == 0:
-                    md_text += "| --- "*len(cells)+"|\n"
-
-            md_text += "\n\n"
-    # with open("details.md","w") as f:
-    #     f.write(md_text)
+    
+    overview = [v for _,v in output.items() if v["assessment"] == "overview"][0]
+    safety = [v for _,v in output.items() if v["assessment"] == "safety"][0]
+    # add overview
+    md_text += f"<details>\n<summary>Overivew</summary>\n\n"
+    md_text += overview["content"] + "\n</details>\n\n"
+    
+    # add performance
+    md_text += f"<details>\n<summary>Performance</summary>\n\n"
+    for title,content in output.items():
+        if content["assessment"] not in  ["overview","safety"]:
+            md_text += f"#### {content['assessment']} - {title}\n\n"
+            md_text += content["content"] + "\n\n"
+    md_text += "</details>\n\n"
+
+    # add safety
+    md_text += f"<details>\n<summary>Safety</summary>\n\n"
+    md_text += safety["content"] + "\n\n" + "</details>\n\n"
+
     return gr.update(value=md_text)
 
 
@@ -239,29 +212,22 @@ def add_article(domain,file,add_to_s3=True, add_to_local=True, file_object=True)
     '''
     if file_object:
         content, _ = read_pdf(file)
-        name = file.name.split("\\")[-1].split(".")[0]
+        filename = file.name.split("\\")[-1]
+        # name = filename.split(".")[0]
     else:
         content = file
-        name = f"temp_{datetime.now().strftime('%Y-%m-%d %H:%M:%S')}"
-
-    abstract,_,end_abstract = extract_key_content(content,["objective","abstract"],["key","words:","methods"],["introduction"])
-    methods,_,end_methods = extract_key_content(content[end_abstract:],["methods"],["results"])
-    if not methods:
-        methods,_,end_methods = extract_key_content(content[end_abstract:],["methods"],["discussion"])
-    results,_,_ = extract_key_content(content[end_methods:],["results"],["discussion"])
+        # filename = file.name
+        filename = f"temp_{datetime.now().strftime('%Y-%m-%d %H:%M:%S')}"
 
     article ={
         "domain":domain,
-        "name":name,
+        "name":filename,
         "content":content,
-        "abstract":abstract,
-        "methods":methods,
-        "results":results,
         "upload_time":datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
     }
     
     if add_to_s3 and file_object:
-        s3_path = upload_fileobj(file,domain,article["name"])
+        s3_path = upload_fileobj(file,domain,filename)
         article["s3_path"] = s3_path
 
     if add_to_local:
@@ -269,7 +235,7 @@ def add_article(domain,file,add_to_s3=True, add_to_local=True, file_object=True)
 
     res = post_item("articles",article)
     if "Error" in res:
-        print(res)
+        print(res["Error"])
         return res
     
     return article
@@ -387,72 +353,31 @@ def remove_device():
 def update_device():
     pass
 
-def process_feedback(text):
-    return text
+# identify article state
+def identify_logic(text):
+    article_logic = [
+        "groups",
+        "levels",
+        "preoperatives"
+    ]
 
-def select_prompts(article,*args):
-    '''
-    select the prompts based on the content and the search terms
-    that was included in the content
+    return {l:l in text.lower() for l in article_logic}
 
-    Parameters
-    ----------
-    article : dict
-        article object
+def select_prompts(text,terms,prompts):
+    selected_templates = set()
+    for t in terms:
+        if all([term in text for term in t["terms"]]):
+            selected_templates.update(t["template_name"])
 
-    Returns
-    -------
-    dict
-        prompts
-    '''
-    
-    # get template names based on the search terms
-    memory = set()
-    prompts = []
-    for t in app_data["terms"]:
-        t["terms"] = t["term"].split(",")
-        if reduce(mul, [s in article["content"] for s in t["terms"]], 1) and t["template_name"] not in memory:
-            # get prompts based from templates
-            template_names = t["template_name"].split(",")
-            for tn in template_names:
-                prompts.extend([p for p in app_data["prompts"] if p["template_name"]==tn])
-                if prompts:
-                    prompts[-1]["prompt"].replace("<--clinical term-->",t["clinical term"])
-                    prompts[-1]["prompt"].replace("<--radiologic term-->",t["clinical term"])
-                    prompts[-1]["prompt"].replace("<--other term-->",t["clinical term"])
-
-            memory.add(t["template_name"])
-
-    # add overview prompts
-    prompts.extend([ov for ov in app_data["prompts"] if ov["assessment_step"]=="overview"])
-    # print("number of prompts",len(prompts))
-    
-    # check if groups, levels and preopratives are in the article
-    article_logic = {}
-    for k,value in logic_keywords.items():
-        article_logic[k] = bool(sum([kw in article["content"] for kw in value]))
-    # print(article_logic)
-
-    # use article_logic to filter prompts
-    prompts = [p for p in prompts 
-                if (p["groups"] == article_logic["groups"] or p["groups"] is None) 
-                and (p["levels"] == article_logic["levels"] or p["levels"] is None) 
-                and (p["preoperatives"] == article_logic["preoperatives"] or p["preoperatives"] is None)]
+    logic = identify_logic(text)
 
-    
-    # print("number of prompts after logic",len(prompts))
-    # early return if no specific result
-    if "".join(args) == "":
-        # print("no args")
-        return prompts
-
-    # # performance metrics and safety metrics filter
-    # for p in prompts:
-    #     if not sum([a in p["clinical term"] for a in args if a]):
-    #         print(p["template_name"])
-    #         prompts.remove(p)
-    # print("number of prompts after args",len(prompts))
-    return prompts
+    selected_prompts = [p for p in prompts if p["template_name"] in selected_templates]
+    overview_prompts = [p for p in prompts if p["assessment_step"] == "overview"]
+    for p in overview_prompts:
+        if all([p[l]==v for l,v in logic.items() if v]):
+            selected_prompts.append(p)
+
+    return selected_prompts
 
 def keyword_search(keywords,full_text):
     keywords_result = {}
@@ -461,4 +386,37 @@ def keyword_search(keywords,full_text):
             keywords_result[k]=list_or([keyword_search(kw,full_text) for kw in k])
         else:
             keywords_result[k]=keyword_search(k,full_text)
-    return keywords_result
+    return keywords_result
+
+def process_prompts(text,prompts):
+    '''
+    process_prompts function receive the text and prompts and return the instruction stream
+
+    Parameters
+    ----------
+    text : str
+        text of the article
+    prompts : list
+        list of prompts
+
+    Returns
+    -------
+    dict
+        processed extraction results from openai api
+    '''
+    res = defaultdict(dict)
+    for p in prompts:
+        inst = [
+            p["prompt"]+", ".join(p["fields"]),
+            p["reformat_inst"]
+        ]
+        inst_stream = create_inst(text,inst)
+        extraction = send_inst(inst_stream)
+
+        res[p["template_name"]] = {
+            "template_name":p["template_name"],
+            "assessment":p["assessment_step"],
+            "content":extraction
+        }
+
+    return res

supplier.py

@@ -1,11 +1,6 @@
 import openai
 from application import *
 
-# import json
-#
-# with open("openai_api_key.json", "r") as f:
-#     openai_api_key = json.load(f)["key"]
-
 openai.api_key = openai_api_key
 
 def execute_prompt(prompt):
@@ -28,3 +23,67 @@ def execute_prompt(prompt):
         presence_penalty=0
     )
     return res.choices[0]["text"] if res.choices else "<error> failed to generate text</error>"
+
+def format(**kwargs):
+    if "format" in kwargs:
+        return kwargs["format"]
+    return kwargs
+
+
+def execute_instruction(article, instruction,model="gpt-3.5-turbo-16k",format="markdown"):
+    '''
+    execute_instruction function takes three arguments: article, instruction and model
+    
+    article: the raw text from the article source
+    instruction: the instruction for the rational execution it needs to complete
+    model: the model used for the rational execution, default to gpt-3.5-turbo-16k
+    format: the format of the table, default to markdown
+
+    return: a string, the result of the rational execution
+    '''
+    msg_stream = [
+        {
+            "role":"system",
+            "content":article
+        }
+    ]
+            
+    msg_stream.append({
+        "role":"user",
+        "content":instruction
+    })
+
+    msg_stream.append({
+        "role":"user",
+        "content":f"Format the table in {format} syntax"
+    })
+
+    res= openai.ChatCompletion.create(
+        model=model,
+        messages=msg_stream,
+        temperature=0)
+    
+    return res["choices"][0]["message"]["content"]
+
+def create_inst(article, instructions):
+    msg_stream = [
+        {
+            "role":"system",
+            "content":article
+        }
+    ]
+    for i in instructions:
+        msg_stream.append({
+            "role":"user",
+            "content":i
+        })
+    
+    return msg_stream
+
+def send_inst(stream, model="gpt-3.5-turbo-16k",temperature=0):
+    res= openai.ChatCompletion.create(
+            model=model,
+            messages=stream,
+            temperature=temperature)
+        
+    return res["choices"][0]["message"]["content"]

ui_study.py

@@ -1,22 +1,16 @@
 import gradio as gr
 
-# from article import *
 from utility import *
 from application import *
 from features import *
 
 def reset():
+    '''
+    reset gradio input and output features in this page.
+    '''
     return (
         gr.Files.update(value=None),
         gr.TextArea.update(value=""),
-        gr.Textbox.update(value=""),
-        gr.Textbox.update(value=""),
-        gr.Textbox.update(value=""),
-        gr.Textbox.update(value=""),
-        gr.Checkbox.update(value=False),
-        gr.Slider.update(value=0),
-        gr.Markdown.update(value=""),
-        gr.Markdown.update(value=""),
         gr.Markdown.update(value="")
     )
 
@@ -25,76 +19,44 @@ with gr.Blocks() as study_page:
         with gr.Column():
             gr.Markdown("## Studies")
             gr.HTML("<hr>")
-    
+
             upload_study = gr.File(label="Upload a clinical study report",type="file")
+            
             input_study = gr.TextArea(label="Or paste a clinical study report content",placeholder="Paste content here...",lines=5)
 
             with gr.Row():
                 btn_reset = gr.Button(value="Reset",variant="stop")
                 btn_add_study = gr.Button(value="Add",variant="primary")
-        with gr.Column():
 
-            performance_metric_1 = gr.Textbox(lines=1, label="identify Key Performance Outcome 1",placeholder="e.g. VAS Score")
-            performance_metric_2 = gr.Textbox(lines=1, label="identify Key Performance Outcome 2",placeholder="e.g. Incidence of Fusion")
-            safety_metric_1 = gr.Textbox(lines=1, label="identify Key Safety Outcome 1",placeholder="e.g. Incidence of Revision")
-            safety_metric_2 = gr.Textbox(lines=1, label="identify Key Safety Outcome 2",placeholder="e.g. Incidence of Nonunion")
-
-            device_options["secondary extraction"] = gr.Checkbox(label="Will a secondary extraction with device stratification be required?",interactive=True)
-            device_options["secondary extraction count"] = gr.Slider(minimum=0, maximum=10, step=1, label="How many secondary extractions with device stratification be required?",interactive=True)
-    
-    gr.Markdown("<hr>")
-    with gr.Row():
         with gr.Column():
-            gr.Markdown("## Literature Report")
-        with gr.Column():
-            bt_refresh = gr.Button(value="Refresh",variant="primary")
-
-    gr.Markdown("<hr>")
-    
-    with gr.Row():
-        
-        with gr.Column(scale=2):
-            overview = gr.Markdown("")
-
-        with gr.Column(scale=3):
-            details = gr.Markdown("")
+            with gr.Row():
+                gr.Markdown("## Literature Report")
+                btn_refresh = gr.Button(value="Refresh",variant="primary")
+            views = gr.Markdown("")
             
-            btn_reset.click(
-                reset,
-                outputs=[
-                    upload_study,
-                    input_study,
-                    performance_metric_1,
-                    performance_metric_2,
-                    safety_metric_1,
-                    safety_metric_2,
-                    device_options["secondary extraction"],
-                    device_options["secondary extraction count"],
-                    overview,
-                    details
-                ]
-            )
+    btn_reset.click(
+        reset,
+        outputs=[
+            upload_study,
+            input_study,
+            views,
+        ]
+    )
 
     btn_add_study.click(
         process_study,
         inputs=[
             upload_study,
             input_study,
-            performance_metric_1,
-            performance_metric_2,
-            safety_metric_1,
-            safety_metric_2
         ],
         outputs=[
-            overview,
-            details
+            views,
         ],
     )
 
-    bt_refresh.click(
+    btn_refresh.click(
         refresh,
         outputs=[
-            overview,
-            details
-        ]
+            views,
+        ],
     )

utility.py

@@ -1,5 +1,4 @@
 import json
-import tempfile
 
 from application import *
 from pdfminer.high_level import extract_text
@@ -143,7 +142,6 @@ def py_dict_to_db_map(py_dict):
             db_map[key] = {"BOOL":value}
         elif value is None:
             db_map[key] = {"NULL":True}
-            
     return db_map
 
 # convert dynamodb list to python list