Spaces:
Running
on
Zero
Running
on
Zero
Y Phung Nguyen
commited on
Commit
·
c8562d7
1
Parent(s):
3b38a6c
Optimise QA round followup extensive
Browse files- .gitignore +2 -1
- pipeline.py +49 -3
- supervisor.py +5 -0
.gitignore
CHANGED
|
@@ -1,3 +1,4 @@
|
|
| 1 |
.env
|
| 2 |
.setup.txt
|
| 3 |
-
__pycache__/
|
|
|
|
|
|
| 1 |
.env
|
| 2 |
.setup.txt
|
| 3 |
+
__pycache__/
|
| 4 |
+
sample.txt
|
pipeline.py
CHANGED
|
@@ -5,6 +5,7 @@ import time
|
|
| 5 |
import logging
|
| 6 |
import threading
|
| 7 |
import concurrent.futures
|
|
|
|
| 8 |
import gradio as gr
|
| 9 |
import spaces
|
| 10 |
from llama_index.core import StorageContext, VectorStoreIndex, load_index_from_storage
|
|
@@ -118,6 +119,46 @@ def _build_refined_query(base_query: str, insights: dict, insights_block: str) -
|
|
| 118 |
return "\n\n".join([section for section in sections if section])
|
| 119 |
|
| 120 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
def _start_clinical_intake_session(session_id: str, plan: dict, base_query: str, original_language: str):
|
| 122 |
questions = plan.get("questions", []) or []
|
| 123 |
if not questions:
|
|
@@ -135,7 +176,8 @@ def _start_clinical_intake_session(session_id: str, plan: dict, base_query: str,
|
|
| 135 |
"answers": [],
|
| 136 |
"decision_reason": plan.get("decision_reason", ""),
|
| 137 |
"initial_hypotheses": plan.get("initial_hypotheses", []),
|
| 138 |
-
"started_at": time.time()
|
|
|
|
| 139 |
}
|
| 140 |
_set_clinical_intake_state(session_id, state)
|
| 141 |
first_prompt = _format_intake_question(
|
|
@@ -144,6 +186,8 @@ def _start_clinical_intake_session(session_id: str, plan: dict, base_query: str,
|
|
| 144 |
max_rounds=max_rounds,
|
| 145 |
target_lang=state["original_language"]
|
| 146 |
)
|
|
|
|
|
|
|
| 147 |
return first_prompt
|
| 148 |
|
| 149 |
|
|
@@ -193,6 +237,8 @@ def _handle_clinical_answer(session_id: str, answer_text: str):
|
|
| 193 |
max_rounds=state["max_rounds"],
|
| 194 |
target_lang=state["original_language"]
|
| 195 |
)
|
|
|
|
|
|
|
| 196 |
return {"type": "question", "prompt": prompt}
|
| 197 |
|
| 198 |
|
|
@@ -235,7 +281,7 @@ def stream_chat(
|
|
| 235 |
def elapsed():
|
| 236 |
return time.time() - session_start
|
| 237 |
|
| 238 |
-
user_id = request.session_hash
|
| 239 |
index_dir = f"./{user_id}_index"
|
| 240 |
has_rag_index = os.path.exists(index_dir)
|
| 241 |
|
|
@@ -285,7 +331,7 @@ def stream_chat(
|
|
| 285 |
if not enable_clinical_intake:
|
| 286 |
_clear_clinical_intake_state(user_id)
|
| 287 |
else:
|
| 288 |
-
intake_state =
|
| 289 |
if intake_state and intake_state.get("awaiting_answer"):
|
| 290 |
logger.info("[INTAKE] Awaiting patient response - processing answer")
|
| 291 |
intake_result = _handle_clinical_answer(user_id, message)
|
|
|
|
| 5 |
import logging
|
| 6 |
import threading
|
| 7 |
import concurrent.futures
|
| 8 |
+
import hashlib
|
| 9 |
import gradio as gr
|
| 10 |
import spaces
|
| 11 |
from llama_index.core import StorageContext, VectorStoreIndex, load_index_from_storage
|
|
|
|
| 119 |
return "\n\n".join([section for section in sections if section])
|
| 120 |
|
| 121 |
|
| 122 |
+
def _hash_prompt_text(text: str) -> str:
|
| 123 |
+
if not text:
|
| 124 |
+
return ""
|
| 125 |
+
digest = hashlib.sha1()
|
| 126 |
+
digest.update(text.strip().encode("utf-8"))
|
| 127 |
+
return digest.hexdigest()
|
| 128 |
+
|
| 129 |
+
|
| 130 |
+
def _extract_pending_intake_prompt(history: list) -> str:
|
| 131 |
+
if not history:
|
| 132 |
+
return ""
|
| 133 |
+
for turn in reversed(history):
|
| 134 |
+
if turn.get("role") != "assistant":
|
| 135 |
+
continue
|
| 136 |
+
content = turn.get("content", "")
|
| 137 |
+
if content.startswith("🩺 Question for clarity"):
|
| 138 |
+
return content
|
| 139 |
+
return ""
|
| 140 |
+
|
| 141 |
+
|
| 142 |
+
def _rehydrate_intake_state(session_id: str, history: list):
|
| 143 |
+
state = _get_clinical_intake_state(session_id)
|
| 144 |
+
if state or not history:
|
| 145 |
+
return state
|
| 146 |
+
pending_prompt = _extract_pending_intake_prompt(history)
|
| 147 |
+
if not pending_prompt:
|
| 148 |
+
return None
|
| 149 |
+
prompt_hash = _hash_prompt_text(pending_prompt)
|
| 150 |
+
if not prompt_hash:
|
| 151 |
+
return None
|
| 152 |
+
with _clinical_intake_lock:
|
| 153 |
+
for existing_id, existing_state in list(_clinical_intake_sessions.items()):
|
| 154 |
+
if existing_state.get("awaiting_answer") and existing_state.get("last_prompt_hash") == prompt_hash:
|
| 155 |
+
if existing_id != session_id:
|
| 156 |
+
_clinical_intake_sessions.pop(existing_id, None)
|
| 157 |
+
_clinical_intake_sessions[session_id] = existing_state
|
| 158 |
+
return existing_state
|
| 159 |
+
return None
|
| 160 |
+
|
| 161 |
+
|
| 162 |
def _start_clinical_intake_session(session_id: str, plan: dict, base_query: str, original_language: str):
|
| 163 |
questions = plan.get("questions", []) or []
|
| 164 |
if not questions:
|
|
|
|
| 176 |
"answers": [],
|
| 177 |
"decision_reason": plan.get("decision_reason", ""),
|
| 178 |
"initial_hypotheses": plan.get("initial_hypotheses", []),
|
| 179 |
+
"started_at": time.time(),
|
| 180 |
+
"last_prompt_hash": ""
|
| 181 |
}
|
| 182 |
_set_clinical_intake_state(session_id, state)
|
| 183 |
first_prompt = _format_intake_question(
|
|
|
|
| 186 |
max_rounds=max_rounds,
|
| 187 |
target_lang=state["original_language"]
|
| 188 |
)
|
| 189 |
+
state["last_prompt_hash"] = _hash_prompt_text(first_prompt)
|
| 190 |
+
_set_clinical_intake_state(session_id, state)
|
| 191 |
return first_prompt
|
| 192 |
|
| 193 |
|
|
|
|
| 237 |
max_rounds=state["max_rounds"],
|
| 238 |
target_lang=state["original_language"]
|
| 239 |
)
|
| 240 |
+
state["last_prompt_hash"] = _hash_prompt_text(prompt)
|
| 241 |
+
_set_clinical_intake_state(session_id, state)
|
| 242 |
return {"type": "question", "prompt": prompt}
|
| 243 |
|
| 244 |
|
|
|
|
| 281 |
def elapsed():
|
| 282 |
return time.time() - session_start
|
| 283 |
|
| 284 |
+
user_id = request.session_hash or "anonymous"
|
| 285 |
index_dir = f"./{user_id}_index"
|
| 286 |
has_rag_index = os.path.exists(index_dir)
|
| 287 |
|
|
|
|
| 331 |
if not enable_clinical_intake:
|
| 332 |
_clear_clinical_intake_state(user_id)
|
| 333 |
else:
|
| 334 |
+
intake_state = _rehydrate_intake_state(user_id, history)
|
| 335 |
if intake_state and intake_state.get("awaiting_answer"):
|
| 336 |
logger.info("[INTAKE] Awaiting patient response - processing answer")
|
| 337 |
intake_result = _handle_clinical_answer(user_id, message)
|
supervisor.py
CHANGED
|
@@ -168,12 +168,17 @@ def _prepare_clinical_question_plan(plan: dict, safe_rounds: int) -> dict:
|
|
| 168 |
if not isinstance(questions, list):
|
| 169 |
questions = []
|
| 170 |
cleaned = []
|
|
|
|
| 171 |
for idx, raw in enumerate(questions):
|
| 172 |
if not isinstance(raw, dict):
|
| 173 |
continue
|
| 174 |
question_text = (raw.get("question") or "").strip()
|
| 175 |
if not question_text:
|
| 176 |
continue
|
|
|
|
|
|
|
|
|
|
|
|
|
| 177 |
entry = dict(raw)
|
| 178 |
entry["question"] = question_text
|
| 179 |
entry["order"] = entry.get("order") or raw.get("id") or (idx + 1)
|
|
|
|
| 168 |
if not isinstance(questions, list):
|
| 169 |
questions = []
|
| 170 |
cleaned = []
|
| 171 |
+
seen = set()
|
| 172 |
for idx, raw in enumerate(questions):
|
| 173 |
if not isinstance(raw, dict):
|
| 174 |
continue
|
| 175 |
question_text = (raw.get("question") or "").strip()
|
| 176 |
if not question_text:
|
| 177 |
continue
|
| 178 |
+
normalized = question_text.lower()
|
| 179 |
+
if normalized in seen:
|
| 180 |
+
continue
|
| 181 |
+
seen.add(normalized)
|
| 182 |
entry = dict(raw)
|
| 183 |
entry["question"] = question_text
|
| 184 |
entry["order"] = entry.get("order") or raw.get("id") or (idx + 1)
|