Spaces:
Sleeping
Sleeping
Update modules/orchestrator.py
Browse files- modules/orchestrator.py +114 -69
modules/orchestrator.py
CHANGED
@@ -1,103 +1,146 @@
|
|
1 |
# modules/orchestrator.py
|
2 |
"""
|
3 |
-
The
|
4 |
-
|
5 |
-
|
|
|
|
|
6 |
"""
|
|
|
7 |
import asyncio
|
8 |
import aiohttp
|
9 |
-
import ast
|
10 |
from itertools import chain
|
11 |
from PIL import Image
|
12 |
|
13 |
-
# Import all our tools
|
14 |
-
from . import gemini_handler, prompts
|
15 |
from .api_clients import (
|
16 |
-
umls_client,
|
17 |
pubmed_client,
|
18 |
clinicaltrials_client,
|
19 |
openfda_client,
|
20 |
rxnorm_client
|
|
|
21 |
)
|
22 |
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
async def run_symptom_synthesis(user_query: str, image_input: Image.Image | None) -> str:
|
42 |
"""The complete, asynchronous pipeline for the Symptom Synthesizer tab."""
|
43 |
if not user_query:
|
44 |
return "Please enter a symptom description or a medical question to begin."
|
45 |
|
46 |
-
# 1
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
except (ValueError, SyntaxError):
|
54 |
-
concepts = [user_query] # Fallback
|
55 |
|
56 |
-
|
|
|
57 |
|
58 |
-
# 2
|
|
|
59 |
async with aiohttp.ClientSession() as session:
|
|
|
60 |
tasks = {
|
61 |
"pubmed": pubmed_client.search_pubmed(session, search_query, max_results=3),
|
62 |
"trials": clinicaltrials_client.find_trials(session, search_query, max_results=3),
|
63 |
-
"openfda": asyncio.gather(*(openfda_client.get_adverse_events(session, c, top_n=3) for c in concepts))
|
64 |
}
|
|
|
65 |
if image_input:
|
66 |
tasks["vision"] = gemini_handler.analyze_image_with_text(
|
67 |
-
"
|
68 |
)
|
69 |
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
# Flatten the list of lists from the OpenFDA gather call
|
78 |
-
fda_results = list(chain.from_iterable(api_data.get('openfda', [])))
|
79 |
-
fda_formatted = _format_data_for_prompt(fda_results, "OpenFDA Adverse Events")
|
80 |
-
|
81 |
-
vision_formatted = api_data.get('vision', "")
|
82 |
-
if isinstance(vision_formatted, Exception):
|
83 |
-
vision_formatted = "Error analyzing image."
|
84 |
|
85 |
-
# 4
|
|
|
86 |
synthesis_prompt = prompts.get_synthesis_prompt(
|
87 |
user_query=user_query,
|
88 |
concepts=concepts,
|
89 |
-
pubmed_data=
|
90 |
-
trials_data=
|
91 |
-
fda_data=
|
92 |
-
vision_analysis=
|
93 |
)
|
94 |
|
95 |
final_report = await gemini_handler.generate_text_response(synthesis_prompt)
|
96 |
|
|
|
|
|
97 |
return f"{prompts.DISCLAIMER}\n\n{final_report}"
|
98 |
|
99 |
|
100 |
-
# ---
|
|
|
101 |
async def run_drug_interaction_analysis(drug_list_str: str) -> str:
|
102 |
"""The complete, asynchronous pipeline for the Drug Interaction Analyzer tab."""
|
103 |
if not drug_list_str:
|
@@ -107,16 +150,16 @@ async def run_drug_interaction_analysis(drug_list_str: str) -> str:
|
|
107 |
if len(drug_names) < 2:
|
108 |
return "Please enter at least two medications to check for interactions."
|
109 |
|
110 |
-
# 1
|
111 |
async with aiohttp.ClientSession() as session:
|
112 |
tasks = {
|
113 |
"interactions": rxnorm_client.run_interaction_check(drug_names),
|
114 |
"safety_profiles": asyncio.gather(*(openfda_client.get_safety_profile(session, name) for name in drug_names))
|
115 |
}
|
116 |
-
|
117 |
-
api_data = dict(zip(tasks.keys(),
|
118 |
|
119 |
-
# 2
|
120 |
interaction_data = api_data.get('interactions', [])
|
121 |
if isinstance(interaction_data, Exception):
|
122 |
interaction_data = [{"error": str(interaction_data)}]
|
@@ -124,20 +167,22 @@ async def run_drug_interaction_analysis(drug_list_str: str) -> str:
|
|
124 |
safety_profiles = api_data.get('safety_profiles', [])
|
125 |
if isinstance(safety_profiles, Exception):
|
126 |
safety_profiles = [{"error": str(safety_profiles)}]
|
127 |
-
|
128 |
-
# Combine safety profiles with their drug names
|
129 |
safety_data_dict = dict(zip(drug_names, safety_profiles))
|
130 |
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
|
|
135 |
synthesis_prompt = prompts.get_drug_interaction_synthesis_prompt(
|
136 |
drug_names=drug_names,
|
137 |
interaction_data=interaction_formatted,
|
138 |
safety_data=safety_formatted
|
139 |
)
|
140 |
-
|
141 |
final_report = await gemini_handler.generate_text_response(synthesis_prompt)
|
142 |
|
|
|
143 |
return f"{prompts.DISCLAIMER}\n\n{final_report}"
|
|
|
1 |
# modules/orchestrator.py
|
2 |
"""
|
3 |
+
The Central Nervous System of Project Asclepius.
|
4 |
+
This module is the master conductor, orchestrating high-performance, asynchronous
|
5 |
+
workflows for each of the application's features. It intelligently sequences
|
6 |
+
calls to API clients and the Gemini handler to transform user queries into
|
7 |
+
comprehensive, synthesized reports.
|
8 |
"""
|
9 |
+
|
10 |
import asyncio
|
11 |
import aiohttp
|
|
|
12 |
from itertools import chain
|
13 |
from PIL import Image
|
14 |
|
15 |
+
# Import all our specialized tools
|
16 |
+
from . import gemini_handler, prompts, utils
|
17 |
from .api_clients import (
|
|
|
18 |
pubmed_client,
|
19 |
clinicaltrials_client,
|
20 |
openfda_client,
|
21 |
rxnorm_client
|
22 |
+
# The umls_client is implicitly used via term extraction, but can be added for deeper analysis
|
23 |
)
|
24 |
|
25 |
+
|
26 |
+
# --- Internal Helper for Data Formatting ---
|
27 |
+
|
28 |
+
def _format_api_data_for_prompt(api_results: dict) -> dict[str, str]:
|
29 |
+
"""
|
30 |
+
Takes the raw dictionary of API results and formats each entry into a
|
31 |
+
clean, readable string suitable for injection into a Gemini prompt.
|
32 |
+
|
33 |
+
Args:
|
34 |
+
api_results (dict): The dictionary of results from asyncio.gather.
|
35 |
+
|
36 |
+
Returns:
|
37 |
+
dict[str, str]: A dictionary with the same keys but formatted string values.
|
38 |
+
"""
|
39 |
+
formatted_strings = {}
|
40 |
+
|
41 |
+
# Format PubMed data
|
42 |
+
pubmed_data = api_results.get('pubmed', [])
|
43 |
+
if isinstance(pubmed_data, list) and pubmed_data:
|
44 |
+
lines = [f"- Title: {a.get('title', 'N/A')} (Journal: {a.get('journal', 'N/A')}, URL: {a.get('url')})" for a in pubmed_data]
|
45 |
+
formatted_strings['pubmed'] = "\n".join(lines)
|
46 |
+
else:
|
47 |
+
formatted_strings['pubmed'] = "No relevant review articles were found on PubMed for this query."
|
48 |
+
|
49 |
+
# Format Clinical Trials data
|
50 |
+
trials_data = api_results.get('trials', [])
|
51 |
+
if isinstance(trials_data, list) and trials_data:
|
52 |
+
lines = [f"- Title: {t.get('title', 'N/A')} (Status: {t.get('status', 'N/A')}, URL: {t.get('url')})" for t in trials_data]
|
53 |
+
formatted_strings['trials'] = "\n".join(lines)
|
54 |
+
else:
|
55 |
+
formatted_strings['trials'] = "No actively recruiting clinical trials were found matching this query."
|
56 |
+
|
57 |
+
# Format OpenFDA Adverse Events data
|
58 |
+
# This data often comes from multiple queries, so we flatten it.
|
59 |
+
fda_data = api_results.get('openfda', [])
|
60 |
+
if isinstance(fda_data, list):
|
61 |
+
# The result is a list of lists, so we flatten it
|
62 |
+
all_events = list(chain.from_iterable(filter(None, fda_data)))
|
63 |
+
if all_events:
|
64 |
+
lines = [f"- {evt['term']} (Reported {evt['count']} times)" for evt in all_events]
|
65 |
+
formatted_strings['openfda'] = "\n".join(lines)
|
66 |
+
else:
|
67 |
+
formatted_strings['openfda'] = "No specific adverse event data was found for this query."
|
68 |
+
else:
|
69 |
+
formatted_strings['openfda'] = "No specific adverse event data was found for this query."
|
70 |
+
|
71 |
+
# Format Vision analysis
|
72 |
+
vision_data = api_results.get('vision', "")
|
73 |
+
if isinstance(vision_data, str) and vision_data:
|
74 |
+
formatted_strings['vision'] = vision_data
|
75 |
+
elif isinstance(vision_data, Exception):
|
76 |
+
formatted_strings['vision'] = f"An error occurred during image analysis: {vision_data}"
|
77 |
+
else:
|
78 |
+
formatted_strings['vision'] = ""
|
79 |
+
|
80 |
+
return formatted_strings
|
81 |
+
|
82 |
+
|
83 |
+
# --- FEATURE 1: Symptom Synthesizer Pipeline ---
|
84 |
+
|
85 |
async def run_symptom_synthesis(user_query: str, image_input: Image.Image | None) -> str:
|
86 |
"""The complete, asynchronous pipeline for the Symptom Synthesizer tab."""
|
87 |
if not user_query:
|
88 |
return "Please enter a symptom description or a medical question to begin."
|
89 |
|
90 |
+
# STEP 1: AI-Powered Concept Extraction
|
91 |
+
# Use Gemini to find the core medical terms in the user's natural language query.
|
92 |
+
term_prompt = prompts.get_term_extraction_prompt(user_query)
|
93 |
+
concepts_str = await gemini_handler.generate_text_response(term_prompt)
|
94 |
+
concepts = utils.safe_literal_eval(concepts_str)
|
95 |
+
if not isinstance(concepts, list) or not concepts:
|
96 |
+
concepts = [user_query] # Fallback to the raw query if parsing fails
|
|
|
|
|
97 |
|
98 |
+
# Use "OR" for a broader, more inclusive search across APIs
|
99 |
+
search_query = " OR ".join(f'"{c}"' for c in concepts)
|
100 |
|
101 |
+
# STEP 2: Massively Parallel Evidence Gathering
|
102 |
+
# Launch all API calls concurrently for maximum performance.
|
103 |
async with aiohttp.ClientSession() as session:
|
104 |
+
# Define the portfolio of data we need to collect
|
105 |
tasks = {
|
106 |
"pubmed": pubmed_client.search_pubmed(session, search_query, max_results=3),
|
107 |
"trials": clinicaltrials_client.find_trials(session, search_query, max_results=3),
|
108 |
+
"openfda": asyncio.gather(*(openfda_client.get_adverse_events(session, c, top_n=3) for c in concepts)),
|
109 |
}
|
110 |
+
# If an image is provided, add the vision analysis to our task portfolio
|
111 |
if image_input:
|
112 |
tasks["vision"] = gemini_handler.analyze_image_with_text(
|
113 |
+
"In the context of the user query, analyze this image objectively. Describe visual features like color, shape, texture, and patterns. Do not diagnose or offer medical advice.", image_input
|
114 |
)
|
115 |
|
116 |
+
# Execute all tasks and wait for them all to complete
|
117 |
+
raw_results = await asyncio.gather(*tasks.values(), return_exceptions=True)
|
118 |
+
api_data = dict(zip(tasks.keys(), raw_results))
|
119 |
+
|
120 |
+
# STEP 3: Data Formatting
|
121 |
+
# Convert the raw JSON/list results into clean, prompt-ready strings.
|
122 |
+
formatted_data = _format_api_data_for_prompt(api_data)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
123 |
|
124 |
+
# STEP 4: The Grand Synthesis
|
125 |
+
# Feed all the structured, evidence-based data into Gemini for the final report generation.
|
126 |
synthesis_prompt = prompts.get_synthesis_prompt(
|
127 |
user_query=user_query,
|
128 |
concepts=concepts,
|
129 |
+
pubmed_data=formatted_data['pubmed'],
|
130 |
+
trials_data=formatted_data['trials'],
|
131 |
+
fda_data=formatted_data['openfda'],
|
132 |
+
vision_analysis=formatted_data['vision']
|
133 |
)
|
134 |
|
135 |
final_report = await gemini_handler.generate_text_response(synthesis_prompt)
|
136 |
|
137 |
+
# STEP 5: Final Delivery
|
138 |
+
# Prepend the mandatory disclaimer to the AI-generated report.
|
139 |
return f"{prompts.DISCLAIMER}\n\n{final_report}"
|
140 |
|
141 |
|
142 |
+
# --- FEATURE 2: Drug Interaction & Safety Analyzer Pipeline ---
|
143 |
+
|
144 |
async def run_drug_interaction_analysis(drug_list_str: str) -> str:
|
145 |
"""The complete, asynchronous pipeline for the Drug Interaction Analyzer tab."""
|
146 |
if not drug_list_str:
|
|
|
150 |
if len(drug_names) < 2:
|
151 |
return "Please enter at least two medications to check for interactions."
|
152 |
|
153 |
+
# STEP 1: Concurrent Drug Data Gathering
|
154 |
async with aiohttp.ClientSession() as session:
|
155 |
tasks = {
|
156 |
"interactions": rxnorm_client.run_interaction_check(drug_names),
|
157 |
"safety_profiles": asyncio.gather(*(openfda_client.get_safety_profile(session, name) for name in drug_names))
|
158 |
}
|
159 |
+
raw_results = await asyncio.gather(*tasks.values(), return_exceptions=True)
|
160 |
+
api_data = dict(zip(tasks.keys(), raw_results))
|
161 |
|
162 |
+
# STEP 2: Data Formatting for AI Synthesis
|
163 |
interaction_data = api_data.get('interactions', [])
|
164 |
if isinstance(interaction_data, Exception):
|
165 |
interaction_data = [{"error": str(interaction_data)}]
|
|
|
167 |
safety_profiles = api_data.get('safety_profiles', [])
|
168 |
if isinstance(safety_profiles, Exception):
|
169 |
safety_profiles = [{"error": str(safety_profiles)}]
|
170 |
+
|
171 |
+
# Combine safety profiles with their drug names for clarity in the prompt
|
172 |
safety_data_dict = dict(zip(drug_names, safety_profiles))
|
173 |
|
174 |
+
# Format the complex data into clean strings
|
175 |
+
interaction_formatted = utils.format_list_as_markdown([str(i) for i in interaction_data]) if interaction_data else "No interactions found."
|
176 |
+
safety_formatted = "\n".join([f"Profile for {drug}: {profile}" for drug, profile in safety_data_dict.items()])
|
177 |
+
|
178 |
+
# STEP 3: AI-Powered Safety Briefing
|
179 |
synthesis_prompt = prompts.get_drug_interaction_synthesis_prompt(
|
180 |
drug_names=drug_names,
|
181 |
interaction_data=interaction_formatted,
|
182 |
safety_data=safety_formatted
|
183 |
)
|
184 |
+
|
185 |
final_report = await gemini_handler.generate_text_response(synthesis_prompt)
|
186 |
|
187 |
+
# STEP 4: Final Delivery
|
188 |
return f"{prompts.DISCLAIMER}\n\n{final_report}"
|