12_voiceAPI

.env

@@ -53,22 +53,22 @@ COHERE_API_KEY="getAone"
 
 
 # ---------- MISTRAL ----------
-MISTRAL_API_KEY="getAone"
-# GENERATION_BACKEND="MISTRAL"   
-# EMBEDDING_BACKEND="MISTRAL"   
-# GENERATION_MODEL_ID="mistral-small-2603"
-# EMBEDDING_MODEL_ID="mistral-embed-2312"
-# EMBEDDING_MODEL_SIZE=1024
-# QDRANT_COLLECTION="1024_docs"
+MISTRAL_API_KEY="BScWRb6OT6xjplE6MJslLWPcLy5BNsjG"
+GENERATION_BACKEND="MISTRAL"   
+EMBEDDING_BACKEND="MISTRAL"   
+GENERATION_MODEL_ID="mistral-small-2603"
+EMBEDDING_MODEL_ID="mistral-embed-2312"
+EMBEDDING_MODEL_SIZE=1024
+QDRANT_COLLECTION="1024_docs"
 
 # ---------- GEMINI ----------
 GEMINI_API_KEY="getAone"
-GENERATION_BACKEND="GEMINI"  
-EMBEDDING_BACKEND="GEMINI" 
-GENERATION_MODEL_ID="gemini-2.5-flash"
-EMBEDDING_MODEL_ID="gemini-embedding-001"
-EMBEDDING_MODEL_SIZE=768
-QDRANT_COLLECTION="768_docs"
+# GENERATION_BACKEND="GEMINI"  
+# EMBEDDING_BACKEND="GEMINI" 
+# GENERATION_MODEL_ID="gemini-2.5-flash"
+# EMBEDDING_MODEL_ID="gemini-embedding-001"
+# EMBEDDING_MODEL_SIZE=768
+# QDRANT_COLLECTION="768_docs"
 
 # ---------- HUGGING FACE ----------
 HF_API_KEY="getAone"

main.py

@@ -4,7 +4,7 @@ from routes.base import base_router
 from routes.assisstant_rag import assisstant_router
 from routes.exam_router import exam_router
 from routes.exam_grading_router import grading_router
-
+from routes.voice_route import voice_router
 app = FastAPI()
 app.add_middleware(
     CORSMiddleware,
@@ -17,4 +17,5 @@ app.add_middleware(
 app.include_router(base_router)
 app.include_router(assisstant_router)
 app.include_router(exam_router)
-app.include_router(grading_router) 
+app.include_router(grading_router) 
+app.include_router(voice_router) 

routes/voice_route.py

@@ -0,0 +1,145 @@
+from fastapi import APIRouter, HTTPException
+from pydantic import BaseModel, Field
+from generation.AssistantRagGenerator import AssistantRagGen
+import json
+import re
+from typing import Any
+
+voice_router = APIRouter(prefix="/voice", tags=["voice"])
+gen = AssistantRagGen()
+
+
+
+class ScoreRequest(BaseModel):
+    userid: str = Field(..., description="Unique identifier of the user")
+    examsessionid: str = Field(..., description="Exam session identifier")
+    transcribedrecording: str = Field(..., min_length=1, description="Raw STT transcript")
+
+class ScoreResponse(BaseModel):
+    userid: str
+    examsessionid: str
+    transcribedrecording: str
+    cheatingscorefrom10: float
+    reasoning: str | None = None
+
+SYSTEM_PROMPT = """
+You are a strict academic integrity analyst for oral exams. You evaluate spoken transcripts for cheating likelihood.
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+STEP 1 — TRANSCRIPT CORRECTION
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+Fix STT (Speech-To-Text) errors only. Rules:
+- Fix misheared words, broken homophones, run-on words
+- Preserve the student's original phrasing and ideas exactly
+- Support Egyptian Arabic (عربي مصري), mixed Arabic-English (Arabizi), and full English
+- Egyptian Arabic examples to correct: "el moshkela" → keep as-is, "bitshتغل" → fix only STT noise
+- Do NOT rephrase, improve, or add ideas
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+STEP 2 — CHEATING SCORE (0.0 to 10.0)
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+Score cheating likelihood. Be STRICT and CALIBRATED.
+
+HIGH SCORE TRIGGERS (each adds 2–4 points):
+- Zero verbal fillers (no "um", "uh", "eh", "يعني", "اممم", "زي", "كده") in a long answer
+- Exact dates, edition numbers, author names, chapter references mid-speech
+- Textbook-perfect definitions with formal academic syntax
+- Abrupt complexity spike — starts casual then suddenly goes academic
+- Perfectly balanced comparisons that read like documentation
+- No personal framing ("I think", "من وجهة نظري", "as far as I know")
+- Answer reads identically to how it would appear in a Wikipedia article or textbook
+
+MEDIUM SCORE TRIGGERS (each adds 1–2 points):
+- Correct but suspiciously complete answer with no hesitation
+- Consistent formal register with no casual speech patterns
+- Named theorem/law/algorithm with full historical attribution
+
+LOW SCORE INDICATORS (reduce score):
+- Natural hesitations ("um", "uh", "يعني", "اممم", "eh")
+- Self-corrections mid-sentence
+- Informal phrasing or personal framing
+- Partial or slightly incorrect information (real recall)
+- Mixed Arabic-English as natural code-switching (not copying)
+
+IMPORTANT RULES:
+- A perfectly clean answer with NO verbal fillers in 3+ sentences = minimum 6.0
+- Wikipedia/textbook-identical phrasing = minimum 7.5
+- Exact citations (author, year, chapter) mid-speech = minimum 8.5
+- Do NOT reward clean answers — real oral exams always have some imperfection
+- Egyptian students code-switching between Arabic and English is NORMAL, do not penalise
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+OUTPUT FORMAT — strict JSON only, no markdown:
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+{
+  "corrected_transcript": "<corrected text, same language as input>",
+  "cheating_score": <float 0.0–10.0>,
+  "reasoning": "<one sentence citing the specific signals that determined the score>"
+}
+""".strip()
+
+
+
+@voice_router.post("/score", response_model=ScoreResponse)
+async def stt_ai_score(request: ScoreRequest) -> ScoreResponse:
+
+    user_message = f"Exam transcript to evaluate:\n\n{request.transcribedrecording}"
+
+    try:
+        raw_output = gen.generator.generate_text(f"{SYSTEM_PROMPT}\n\n{user_message}")
+    except Exception as e:
+        raise HTTPException(status_code=502, detail=f"Generation backend error: {e}")
+
+    try:
+        parsed = extract_json_payload(raw_output)
+
+        corrected_transcript: str = parsed["corrected_transcript"]
+        cheating_score: float     = float(parsed["cheating_score"])
+        cheating_score            = max(0.0, min(10.0, cheating_score))
+
+    except (KeyError, ValueError) as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"Failed to parse model response: {e}. Raw: {str(raw_output)[:300]}",
+        )
+
+    return ScoreResponse(
+    userid=request.userid,
+    examsessionid=request.examsessionid,
+    transcribedrecording=corrected_transcript,
+    cheatingscorefrom10=cheating_score,
+    reasoning=parsed.get("reasoning"),
+)
+
+
+
+
+
+
+
+def extract_json_payload(raw: Any) -> dict:
+    if isinstance(raw, dict):
+        if "corrected_transcript" in raw:
+            return raw
+        for val in raw.values():
+            try:
+                return extract_json_payload(val)
+            except ValueError:
+                continue
+
+    if isinstance(raw, str):
+        clean = re.sub(r"```(?:json)?|```", "", raw).strip()
+        try:
+            parsed = json.loads(clean)
+            return extract_json_payload(parsed)   # recurse on parsed result
+        except json.JSONDecodeError:
+            pass
+
+    if isinstance(raw, list):
+        for item in raw:
+            try:
+                return extract_json_payload(item)
+            except ValueError:
+                continue
+
+    raise ValueError(f"Could not find target payload in: {str(raw)[:200]}")