File size: 9,139 Bytes
4d67ec7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 | """
pipeline.py — Main RAG orchestrator
Ties together:
1. Retriever → get top-k NCERT chunks from Supabase
2. PromptBuilder → construct grounded LLM prompts
3. Generator → call Groq to answer / justify / verify
4. Confidence → produce a single confidence score
Usage:
from rag.pipeline import run_pipeline
result = run_pipeline("Which organelle is the powerhouse of the cell?")
print(result)
"""
from rag.retriever import retrieve_chunks
from rag.prompt_builder import (
build_answer_prompt,
build_justification_prompt,
build_verifier_prompt,
)
from rag.generator import generate_answer, generate_justification, verify_answer_llm
from rag.confidence import compute_confidence
from Classification.classify_and_extract import classify_question
def run_pipeline(
question: str,
chapter_filter: str = None,
top_k: int = 5,
use_classifier: bool = False,
) -> dict:
"""
End-to-end RAG pipeline for a NEET Biology question.
Args:
question : The student's question string.
chapter_filter : Optional chapter name to restrict retrieval scope.
Auto-populated by BERT classifier if use_classifier=True.
top_k : Number of NCERT chunks to retrieve.
use_classifier : If True, auto-classify question to get chapter + topics.
Returns a structured dict:
{
"question" : str,
"classification" : {chapter, chapter_id, chapter_confidence, topics},
"answer" : str,
"justification" : list[str],
"sources" : list[{chapter, section, section_title, text_content, similarity}],
"confidence" : float,
"verified" : bool,
"verifier_note" : str,
}
"""
# ── Step 0: BERT Classification + Topic Extraction ──────────────────────
classification = None
if use_classifier:
print("[Pipeline] Step 0: BERT classification + topic extraction...")
classification = classify_question(question)
predicted_chapter = classification["chapter"]
print(f"[Pipeline] Predicted chapter: {predicted_chapter} (Not used as filter)")
# ── Confidence gate: reject if BERT is too uncertain or OOD ──
if classification.get("rejected") or classification["chapter_confidence"] < 0.45:
reject_reason = classification.get("rejection_reason") or f"BERT confidence ({classification['chapter_confidence']:.2%}) is below the 45% threshold."
print(f"[Pipeline] ⚠️ Classifier Rejected — {reject_reason}")
return {
"question" : question,
"classification" : classification,
"answer" : "This question does not appear to be related to NEET Biology.",
"justification" : [],
"sources" : [],
"confidence" : 0.0,
"verified" : False,
"verifier_note" : reject_reason,
"rejected" : True,
}
# ── Step 1: Retrieve relevant NCERT chunks ──────────────────────────────
print("[Pipeline] Step 1: Retrieving NCERT chunks (Global DB Search)...")
chunks = retrieve_chunks(question, chapter_filter=None, top_k=top_k)
if not chunks:
return {
"question" : question,
"classification" : classification,
"answer" : "Insufficient context: no relevant NCERT passages found.",
"justification" : [],
"sources" : [],
"confidence" : 0.0,
"verified" : False,
"verifier_note" : "No chunks retrieved — cannot verify.",
}
# Extract top chapter/section for prompt context
top_chunk = chunks[0]
top_chapter = top_chunk.get("chapter", "Unknown Chapter")
top_section = top_chunk.get("section_title", "Unknown Topic")
# ── Step 3: Generate answer + justification in a single Groq call ──────
print("[Pipeline] Step 3: Generating answer + justification via Groq...")
answer_msgs = build_answer_prompt(question, top_chapter, top_section, chunks)
# Append instruction to also include justification in the same response
answer_msgs[-1]["content"] += (
"\n\nAfter the answer, on a new line write '---JUSTIFICATION---' "
"and then provide a numbered step-by-step biological justification "
"for why the answer is correct, citing only the NCERT context above."
)
raw_response = generate_answer(answer_msgs)
# Parse answer and justification apart
if "---JUSTIFICATION---" in raw_response:
parts = raw_response.split("---JUSTIFICATION---", 1)
answer = parts[0].strip()
just_raw = parts[1].strip()
else:
answer = raw_response.strip()
just_raw = ""
# Parse justification into numbered steps
if just_raw:
lines = just_raw.strip().split("\n")
justification = [l.strip() for l in lines if l.strip() and (
l.strip()[0].isdigit() or l.strip().lower().startswith("step")
)]
if not justification:
justification = [l.strip() for l in lines if l.strip()]
else:
justification = []
# ── Step 5: Verify answer (LLM Judge) ───────────────────────────────────
print("[Pipeline] Step 5: Verifying answer via LLM Judge...")
verif_msgs = build_verifier_prompt(question, answer, chunks)
verif_result = verify_answer_llm(verif_msgs)
verified = verif_result["verified"]
verifier_note = verif_result["explanation"]
# ── Step 6: Compute confidence ──────────────────────────────────────────
retrieval_scores = [c.get("similarity", 0.0) for c in chunks]
confidence = compute_confidence(retrieval_scores, verified)
# ── Step 7: Build sources list (no raw embeddings) ─────────────────────
sources = [
{
"chapter" : c.get("chapter"),
"section" : c.get("section"),
"section_title": c.get("section_title"),
"text_content" : c.get("text_content"),
"similarity" : round(c.get("similarity", 0.0), 4),
}
for c in chunks
]
return {
"question" : question,
"classification" : classification,
"answer" : answer,
"justification" : justification,
"sources" : sources,
"confidence" : confidence,
"verified" : verified,
"verifier_note" : verifier_note,
}
if __name__ == "__main__":
import json, sys
# ── Chapters currently in your Supabase DB ──────────────────────────────
# Kebo101 → The Living World
# Kebo109 → Biomolecules
TEST_QUESTIONS = [
"What is the role of cofactors in enzyme activity?", # Biomolecules
"What are primary and secondary metabolites in living organisms?", # Biomolecules
"How does carbonic anhydrase speed up chemical reactions?", # Biomolecules
"What is the difference between metabolism, anabolism and catabolism?", # Biomolecules
"What is binomial nomenclature and who proposed it?", # The Living World
]
question = TEST_QUESTIONS[0] # ← change index (0-4) to test different questions
debug = "--debug" in sys.argv # run with: python -m rag.pipeline --debug
print(f"\n{'='*60}")
print(f"QUESTION: {question}")
print(f"{'='*60}\n")
result = run_pipeline(question)
if debug:
# Show full retrieved context so you can verify what the LLM saw
print("\n── RETRIEVED CONTEXT ──────────────────────────────────────")
for i, src in enumerate(result["sources"], 1):
print(f"\n[{i}] ({src['similarity']:.3f}) {src['chapter']} > {src['section_title']}")
print(f" {src['text_content'][:300]}...")
print()
print("── ANSWER ─────────────────────────────────────────────────")
print(result["answer"])
print("\n── JUSTIFICATION ──────────────────────────────────────────")
for step in result["justification"]:
print(f" {step}")
print(f"\n── CONFIDENCE: {result['confidence']} | VERIFIED: {result['verified']} ──")
if result.get("verifier_note"):
print(f" Verifier note: {result['verifier_note']}")
|