BlackBox / src /evaluation /evaluate.py
AbdullahKhanSherwani's picture
Final updates
4afbbc2 verified
"""
Evaluation module with Faithfulness and Relevancy metrics.
Faithfulness: claim extraction → verification → % supported by context.
Relevancy: alternate query generation → cosine similarity with original query.
Uses GPT (openai/gpt-oss-120b via NVIDIA API) for both generation and LLM-as-judge.
Runs in parallel batches of 10 jobs (10 × 3 LLM calls = 30 RPM, 62s window).
"""
import json
import os
import re
import sys
import hashlib
import time
import threading
from concurrent.futures import ThreadPoolExecutor, as_completed
import numpy as np
import pandas as pd
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))))
from src.llm.client import call_llm, call_eval_llm, MODEL_GPT
from src.retrieval.query import load_model, init_pinecone, retrieve, available_strategies
from src.retrieval.hybrid import build_bm25_index, load_reranker, hybrid_retrieve, bm25_retrieve
from src.generation.generate import generate_answer, build_prompt
BASE_DIR = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
CACHE_PATH = os.path.join(BASE_DIR, "data", "processed", "eval_llm_cache.json")
# LLM used for both answer generation and LLM-as-judge evaluation
EVAL_MODEL = MODEL_GPT
# Batch processing constants (30 RPM rate limit → 10 jobs × 3 LLM calls = 30/batch)
EVAL_BATCH_SIZE = 10
EVAL_WORKERS = 10
BATCH_WINDOW_S = 62 # seconds; sleep out the remainder after each batch
# Thread-safety locks
_cache_lock = threading.Lock()
_csv_lock = threading.Lock()
# Fixed test set: Q2, Q3, Q4, Q6, Q9 from MANUAL_COMPARE_QA (specific NTSB queries
# with known reference answers) + 7 corpus-grounded queries = 12 total (within 10-20).
# All general queries are anchored to named reports so retrieval quality is measurable.
EVAL_QUERIES = [
# Hard specific queries — require precise detail extraction from reports
"Based strictly on AAR-14/01 for Asiana Airlines Flight 214, what were the exact jumpseat designators for the two flight attendants ejected onto the runway?",
"Based strictly on AAR-14/01 for Asiana Airlines Flight 214, exactly how many total flight hours did the PF have in the Boeing 777 prior to the accident flight?",
"Based strictly on AAR-00/03 for TWA Flight 800, which debris field was the smallest and what were the exact fuselage station markers for the wreckage it contained?",
"Based strictly on AAR-01/01 for United Airlines Flight 585, exactly how many flight hours and minutes of PIC experience did the captain have specifically in the Boeing 737-200 prior to the accident?",
"Based strictly on AAR-01/02 for American Airlines Flight 1420, exactly how many feet to the left of the centerline were the main landing gear tire marks at the end of runway 4R?",
"Based strictly on AAR-00/02 for Federal Express Flight 14, exactly how many minutes elapsed from the Condition One alarm until five ARFF vehicles were actively engaged in fire suppression?",
"Based strictly on AAR-18/01 for American Airlines Flight 383, exactly how many seconds elapsed between the start of the right engine failure and the time the spar valve was closed?",
# Cross-report synthesis queries — require reasoning across multiple documents
"Across the Korean Air Flight 801 and Asiana Airlines Flight 214 NTSB reports, what specific crew training deficiencies and simulator limitations were identified as contributing factors?",
"Comparing the NTSB reports for TWA Flight 800 and United Airlines Flight 585, what differences in the post-accident investigation methodology and evidence recovery processes were documented?",
"What were the specific autopilot and autothrottle anomalies documented in the Asiana Airlines Flight 214 NTSB report, and how did they interact with crew workload during the approach?",
"What were the flight hours of the first officer in the national air crash afghanistan?",
"Based strictly on AAR-00/01 for Korean Air Flight 801, what was the exact decision height for the ILS approach at Guam and at what altitude did the crew first receive a GPWS warning?",
"Based strictly on AAR-14/01 for Asiana Airlines Flight 214, state the exact CVR timestamp in hours, minutes, and seconds when the stick shaker first activated, the exact indicated airspeed at that moment, and the exact radio altitude recorded simultaneously by the FDR.",
]
MANUAL_COMPARE_QA = [
{
"id": "Q1",
"question": "Based strictly on the NTSB Aviation Accident Report AAR-18/01 for American Airlines Flight 383, provide the exact chronological telemetry from the Flight Data Recorder (FDR): Exactly how many seconds elapsed between the start of the right engine failure and the time the spar valve was finally closed to cut off the fuel supply?",
"reference_answer": "Answer: Exactly 164 seconds elapsed between the start of the right engine failure and the closure of the spar valve.",
},
{
"id": "Q2",
"question": "Based strictly on the NTSB Aircraft Accident Report AAR-14/01 for Asiana Airlines Flight 214, detail the survival and evacuation factors: When the tail section of the aircraft struck the seawall and separated from the fuselage, two flight attendants seated in the aft cabin were ejected onto the runway and survived. What were the exact jumpseat designators for these two flight attendants?",
"reference_answer": "The two ejected flight attendants were seated in jumpseats 4L and 4R.",
},
{
"id": "Q3",
"question": "Based strictly on the NTSB Aircraft Accident Report AAR-14/01 for Asiana Airlines Flight 214, one of the fatally injured passengers was located outside the aircraft and was struck by two ARFF vehicles. According to the NTSB, what specific substance visually obscured this passenger from the view of the ARFF drivers?",
"reference_answer": "The passenger was visually obscured by Aqueous Film-Forming Foam (AFFF), also known as firefighting foam.",
},
{
"id": "Q4",
"question": "Based strictly on the NTSB Aircraft Accident Report AAR-14/01 for Asiana Airlines Flight 214, exactly how many flight hours of experience did the PM have acting specifically as an Instructor Pilot in the Boeing 777 prior to this accident flight?",
"reference_answer": "The PM had 0 hours of experience acting specifically as an Instructor Pilot in the Boeing 777 prior to the accident.",
},
{
"id": "Q5",
"question": "Based strictly on the NTSB Aircraft Accident Report AAR-14/01 for Asiana Airlines Flight 214, exactly how many total flight hours did the PF have in the Boeing 777 prior to the accident flight?",
"reference_answer": "The PF had a total of 43 hours in the Boeing 777 prior to the accident flight.",
},
{
"id": "Q6",
"question": "Based strictly on the NTSB Aircraft Accident Report AAR-00/03 for Trans World Airlines (TWA) Flight 800, wreckage was recovered from three debris fields (red, yellow, and green zones). Which field was the smallest, was contained within the red zone on its northeastern side, and what were the exact fuselage station markers for the wreckage it contained?",
"reference_answer": "The Yellow Zone was the smallest, and it contained wreckage from STA 840 to STA 1000.",
},
{
"id": "Q7",
"question": "Based strictly on the NTSB Aircraft Accident Report AAR-01/01 for United Airlines Flight 585, the captain had 9,902 total flight hours with United. Exactly how many flight hours and minutes of PIC experience did he have specifically in the Boeing 737-200 prior to the accident flight?",
"reference_answer": "The captain had 167 hours and 17 minutes of PIC experience specifically in the Boeing 737-200.",
},
{
"id": "Q8",
"question": "Based strictly on the NTSB Aircraft Accident Report AAR-01/02 for American Airlines Flight 1420, exactly how many feet to the left of the centerline were the main landing gear tire marks located at the end of runway 4R?",
"reference_answer": "The tire marks were located 14 feet to the left of the runway centerline at the end of the runway surface.",
},
{
"id": "Q9",
"question": "Based strictly on the NTSB Aircraft Accident Report AAR-00/01 for Korean Air Flight 801, what was the exact diameter of the severed fuel oil pipeline, and approximately how many gallons of oil were spilled?",
"reference_answer": "The pipeline was 12 inches in diameter, and approximately 1,500 gallons of oil were spilled.",
},
{
"id": "Q10",
"question": "Based strictly on the NTSB Aircraft Accident Report AAR-00/02 for Federal Express Flight 14, according to the ARFF fire crew chief, exactly how many minutes elapsed from the Condition One alarm until five ARFF vehicles were actively engaged in fire suppression at the accident site?",
"reference_answer": "Exactly 4 minutes elapsed from the time of the alarm until five ARFF vehicles were actively engaged in fire suppression.",
},
]
def _parse_json(text):
"""Parse JSON from an LLM response, handling markdown code blocks."""
if text.startswith("```"):
text = text.split("\n", 1)[1].rsplit("```", 1)[0].strip()
return json.loads(text)
def _normalize_verified_claims(raw):
"""Normalize judge output into list[dict(claim,supported,reasoning)].
Handles common malformed shapes from LLM outputs:
- JSON string that itself contains JSON
- single dict
- list of strings (claims only)
- mixed lists
"""
if raw is None:
return []
# If model returned JSON as a string value, parse one more time.
if isinstance(raw, str):
try:
raw = _parse_json(raw)
except Exception:
return []
if isinstance(raw, dict):
raw = [raw]
if not isinstance(raw, list):
return []
out = []
for item in raw:
if isinstance(item, dict):
claim = str(item.get("claim", "")).strip()
reasoning = str(item.get("reasoning", "")).strip()
supported = item.get("supported", False)
supported = bool(supported) if isinstance(supported, (bool, int)) else str(supported).strip().lower() in {"true", "yes", "1"}
if claim:
out.append({"claim": claim, "supported": supported, "reasoning": reasoning})
elif isinstance(item, str):
claim = item.strip()
if claim:
out.append({"claim": claim, "supported": False, "reasoning": "string_only_claim"})
return out
def _normalize_alternates(raw, n=3):
"""Normalize alternate-query outputs into a clean list of strings."""
if raw is None:
return []
if isinstance(raw, str):
try:
raw = _parse_json(raw)
except Exception:
return []
if isinstance(raw, dict):
raw = raw.get("alternates", [])
if not isinstance(raw, list):
return []
out = []
for item in raw:
if isinstance(item, str):
s = item.strip()
elif isinstance(item, dict):
s = str(item.get("query", item.get("text", ""))).strip()
else:
s = str(item).strip()
if s:
out.append(s)
# De-duplicate while preserving order.
deduped = []
seen = set()
for s in out:
k = s.lower()
if k in seen:
continue
seen.add(k)
deduped.append(s)
return deduped[:n]
def _load_cache(cache_path=CACHE_PATH):
"""Load persistent evaluation cache from disk."""
if os.path.exists(cache_path):
try:
with open(cache_path, "r", encoding="utf-8") as f:
return json.load(f)
except json.JSONDecodeError:
return {}
return {}
def _save_cache(cache, cache_path=CACHE_PATH):
"""Persist evaluation cache to disk."""
os.makedirs(os.path.dirname(cache_path), exist_ok=True)
with open(cache_path, "w", encoding="utf-8") as f:
json.dump(cache, f)
def _cached_llm(prompt, system=None, cache=None):
"""Call LLM (GPT) with prompt-level memoization. Thread-safe via _cache_lock."""
if cache is None:
return call_llm(prompt, system=system, model=EVAL_MODEL)
key_src = json.dumps({"system": system or "", "prompt": prompt}, ensure_ascii=False, sort_keys=True)
key = hashlib.sha256(key_src.encode("utf-8")).hexdigest()
with _cache_lock:
if key in cache:
return cache[key]
out = call_llm(prompt, system=system, model=EVAL_MODEL)
with _cache_lock:
cache[key] = out
return out
# ── Context builder ───────────────────────────────────────────────────────────
def build_faithfulness_context(query, matches):
"""Return the exact context string the LLM saw during generation.
Calls build_prompt() so the faithfulness judge verifies claims against
the same executive summaries, metadata headers, contextualized_text,
and context_summary fields that were passed to the generation LLM —
not just the stripped raw chunk text.
Public so app.py can import and use it for UI faithfulness scoring.
"""
_, user_prompt = build_prompt(query, matches)
# user_prompt structure: "--- Context ---\n{context}\n\n--- Question ---\n{query}"
# Extract only the context section so the judge isn't confused by the question.
parts = user_prompt.split("\n\n--- Question ---\n", 1)
context_section = parts[0].replace("--- Context ---\n", "", 1)
return context_section
# ── Faithfulness ──────────────────────────────────────────────────────────────
def extract_and_verify_claims(answer, context_texts, query=None, cache=None):
"""Single-call faithfulness judge: extract claims and verify support.
Returns list of dicts with keys: claim, supported (bool), reasoning.
"""
context_str = "\n\n".join(context_texts)
question_block = f"\n===Question===\n{query}\n" if query else ""
prompt = f"""You are an impartial fact-checker for aviation accident analysis.
Your task is to extract factual claims from the answer and verify each claim against the context.
{question_block}
===Context===
{context_str}
===Answer===
{answer}
===Return Requirements===
1. Return a JSON array only.
2. Each element must have exactly keys: "claim", "supported", "reasoning".
3. "claim": one atomic factual claim extracted from the answer.
4. "supported": true only if directly supported by context or by the Question, else false.
5. "reasoning": one short sentence citing context support or mismatch.
6. Include all meaningful POSITIVE factual claims from the answer.
7. For factual entities (report identifiers, flight numbers, dates, aircraft types) that appear
verbatim in the Question, treat the Question as an authoritative source when verifying those claims.
8. SKIP any claim that asserts absence of information (e.g. "the report does not mention X",
"no information is provided about Y", "the report only documents Z and nothing else").
Absence claims cannot be verified from context and must not be scored.
9. SKIP any claim that is purely a summary or connector phrase with no standalone fact.
10. No text outside JSON.
Example:
[{{"claim": "...", "supported": true, "reasoning": "..."}}]"""
try:
raw = _parse_json(_cached_llm(prompt, cache=cache))
except (json.JSONDecodeError, ValueError):
return []
return _normalize_verified_claims(raw)
def compute_faithfulness(answer, context_texts, query=None, cache=None):
"""Compute faithfulness score: fraction of claims supported by context.
Returns (score, details) where score is 0.0-1.0 and details is the
list of verified claims.
"""
verified = extract_and_verify_claims(answer, context_texts, query=query, cache=cache)
if not verified:
return 1.0, []
supported = sum(1 for v in verified if v.get("supported", False))
score = supported / len(verified) if verified else 0.0
return score, verified
# ── Relevancy ─────────────────────────────────────────────────────────────────
def generate_alternate_queries(query, n=3):
"""Generate n alternate phrasings of the query using the LLM (GPT)."""
prompt = f"""You are an aviation safety research assistant.
Your task is to generate {n} alternative phrasings of the following question about aviation accidents.
===Question===
{query}
===Return Requirements===
1. Each alternative must capture the same information need but use different wording.
2. Maintain aviation domain terminology where appropriate.
3. The output must be a properly formatted JSON array of exactly {n} strings.
4. Do NOT include any additional explanation, commentary, or text outside of the JSON array.
5. The output must include only the JSON array and no additional text before or after it.
Example: ["alt 1", "alt 2", "alt 3"]"""
try:
return _parse_json(call_llm(prompt, model=EVAL_MODEL))
except (json.JSONDecodeError, ValueError):
return []
def _extract_final_answer(answer):
"""Extract only the Answer section from a structured LLM response.
The generation prompt enforces an 'Answer:' section at the end.
Falls back to the full answer if the section is not found.
"""
match = re.search(r"(?i)answer\s*[:\-]?\s*(.*)", answer, re.DOTALL)
if match:
return match.group(1).strip()
return answer
def compute_relevancy(query, answer, jina_model, cache=None):
"""Compute relevancy by generating 3 questions from the final answer, then
computing cosine similarity between those questions and the original query.
Returns (score, alternates) where score is mean cosine similarity (0-1).
"""
final_answer = _extract_final_answer(answer)
prompt = f"""You are an aviation safety research assistant.
Given the answer below, generate exactly 3 questions that this answer directly addresses.
===Answer===
{final_answer}
===Return Requirements===
1. Output a JSON array of exactly 3 strings.
2. Each string must be a question that the answer above can answer.
3. Use aviation domain terminology where appropriate.
4. No text outside the JSON array.
Example: ["question 1?", "question 2?", "question 3?"]"""
try:
raw_alternates = _parse_json(_cached_llm(prompt, cache=cache))
except (json.JSONDecodeError, ValueError):
raw_alternates = []
alternates = _normalize_alternates(raw_alternates, n=3)
if not alternates:
return 0.0, []
all_texts = [query] + alternates
embeddings = jina_model.encode(texts=all_texts, task="text-matching")
embeddings = np.array(embeddings)
query_emb = embeddings[0]
alt_embs = embeddings[1:]
similarities = []
for alt_emb in alt_embs:
cos_sim = np.dot(query_emb, alt_emb) / (
np.linalg.norm(query_emb) * np.linalg.norm(alt_emb) + 1e-10
)
similarities.append(float(cos_sim))
return float(np.mean(similarities)), alternates
# ── Retrieval metrics ─────────────────────────────────────────────────────────
def eval_retrieval(results):
"""Compute retrieval quality metrics from results (Pinecone matches or dicts)."""
if not results:
return {"avg_score": 0, "min_score": 0, "max_score": 0, "num_unique_reports": 0}
scores = []
ntsb_nos = set()
for r in results:
if hasattr(r, "score"):
scores.append(r.score)
else:
scores.append(r.get("score", 0))
ntsb = r.metadata.get("ntsb_no", "") if hasattr(r, "metadata") else r.get("ntsb_no", "")
if ntsb:
ntsb_nos.add(ntsb)
return {
"avg_score": sum(scores) / len(scores),
"min_score": min(scores),
"max_score": max(scores),
"num_unique_reports": len(ntsb_nos),
}
# ── Main evaluation loop ─────────────────────────────────────────────────────
def _load_completed(output_path):
"""Load already-completed (query, strategy, mode) tuples from a CSV."""
completed = set()
if os.path.exists(output_path):
df = pd.read_csv(output_path, on_bad_lines='skip')
for _, row in df.iterrows():
completed.add((row["query"], row["strategy"], row["mode"]))
return completed
def _append_result(result, output_path):
"""Append a single result dict to the CSV. Thread-safe via _csv_lock."""
row = {k: v for k, v in result.items() if k not in ("faith_details", "rel_alternates")}
df = pd.DataFrame([row])
with _csv_lock:
write_header = not os.path.exists(output_path)
df.to_csv(output_path, mode="a", header=write_header, index=False)
def _eval_single_job(args):
"""Execute one (query, strategy, mode) evaluation job.
Runs retrieve → generate (GPT) → faithfulness → relevancy and records
retrieval_time, generation_time, and total_time for §3D report metrics.
Thread-safe: uses module-level _cache_lock for shared cache access.
"""
query = args["query"]
strategy = args["strategy"]
mode = args["mode"]
jina_model = args["jina_model"]
index = args["index"]
bm25_cache = args["bm25_cache"]
reranker = args["reranker"]
top_k = args["top_k"]
compute_faith = args["compute_faith"]
compute_rel = args["compute_rel"]
cache = args["cache"]
use_hyde = args["use_hyde"]
allow_bm25_fallback = args["allow_bm25_fallback"]
t_total = time.perf_counter()
# ── Retrieve ────────────────────────────────────────────────────────────────
t_ret = time.perf_counter()
try:
if mode == "hybrid" and bm25_cache and strategy in bm25_cache and reranker:
bm25, chunks = bm25_cache[strategy]
matches = hybrid_retrieve(
query, strategy, top_k=top_k,
model=jina_model, index=index,
bm25=bm25, chunks=chunks, reranker=reranker,
use_hyde=use_hyde,
)
else:
matches = retrieve(query, strategy, top_k=top_k, model=jina_model, index=index)
except Exception as e:
if allow_bm25_fallback and bm25_cache and strategy in bm25_cache:
print(f" Warning: Pinecone failed [{mode}/{strategy}] -> BM25 fallback. {e}")
bm25, chunks = bm25_cache[strategy]
matches = bm25_retrieve(query, bm25, chunks, top_k=top_k)
else:
raise
retrieval_time = round(time.perf_counter() - t_ret, 3)
# Build the full context string the LLM will see (executive summaries +
# metadata headers + contextualized_text). Used for both generation and
# faithfulness verification so the judge sees exactly what the LLM saw.
faithfulness_context = build_faithfulness_context(query, matches)
# ── Generate (GPT) ──────────────────────────────────────────────────────────
t_gen = time.perf_counter()
try:
answer = generate_answer(query, matches, llm_provider="gpt")
except Exception as e:
print(f" Warning: generation failed [{mode}/{strategy}] -> placeholder. {e}")
answer = "Generation unavailable due to LLM connection error."
generation_time = round(time.perf_counter() - t_gen, 3)
total_time = round(time.perf_counter() - t_total, 3)
ret_metrics = eval_retrieval(matches)
# ── Faithfulness (GPT-as-judge) ──────────────────────────────────────────────
if compute_faith:
try:
faith_score, faith_details = compute_faithfulness(
answer, [faithfulness_context], query=query, cache=cache
)
except Exception as e:
print(f" Warning: faithfulness failed [{mode}/{strategy}] -> 0.0. {e}")
faith_score, faith_details = 0.0, []
else:
faith_score, faith_details = 0.0, []
# ── Relevancy (GPT-as-judge + Jina embeddings) ───────────────────────────────
if compute_rel:
try:
rel_score, rel_alternates = compute_relevancy(query, answer, jina_model, cache=cache)
except Exception as e:
print(f" Warning: relevancy failed [{mode}/{strategy}] -> 0.0. {e}")
rel_score, rel_alternates = 0.0, []
else:
rel_score, rel_alternates = 0.0, []
return {
"query": query,
"strategy": strategy,
"mode": mode,
"answer": answer,
"num_chunks": len(matches),
**ret_metrics,
"retrieval_time": retrieval_time,
"generation_time": generation_time,
"total_time": total_time,
"faithfulness": round(faith_score, 3),
"relevancy": round(rel_score, 3),
"faith_details": faith_details,
"rel_alternates": rel_alternates,
}
def run_evaluation(queries, strategies, jina_model, index, mode="semantic",
bm25_cache=None, reranker=None, output_path=None,
top_k=15, compute_faith=True, compute_rel=True,
cache=None, use_hyde=False, allow_bm25_fallback=False,
workers=EVAL_WORKERS, batch_size=EVAL_BATCH_SIZE):
"""Run evaluation across all queries and strategies in parallel batches.
Batches of `batch_size` jobs run with `workers` threads. After each batch
(except the last) we sleep out the remainder of a 62-second window to stay
within the 30 RPM rate limit (10 jobs × 3 LLM calls = 30 calls/batch).
mode: "semantic" or "hybrid"
output_path: incremental CSV save with resume support.
Returns list of result dicts.
"""
completed = _load_completed(output_path) if output_path else set()
jobs = []
for query in queries:
for strategy in strategies:
if (query, strategy, mode) in completed:
print(f" Skipping (done): [{mode}/{strategy}] {query[:50]}...")
continue
jobs.append({
"query": query, "strategy": strategy, "mode": mode,
"jina_model": jina_model, "index": index,
"bm25_cache": bm25_cache, "reranker": reranker,
"top_k": top_k, "compute_faith": compute_faith,
"compute_rel": compute_rel, "cache": cache,
"use_hyde": use_hyde, "allow_bm25_fallback": allow_bm25_fallback,
})
if not jobs:
print(" All entries already completed.")
return []
total = len(jobs)
results = []
done = 0
n_batches = (total + batch_size - 1) // batch_size
for batch_idx, batch_start in enumerate(range(0, total, batch_size), 1):
batch = jobs[batch_start:batch_start + batch_size]
is_last = batch_idx == n_batches
t_batch = time.perf_counter()
print(f"\n Batch {batch_idx}/{n_batches}: "
f"jobs {batch_start+1}{batch_start+len(batch)} of {total}")
with ThreadPoolExecutor(max_workers=min(workers, len(batch))) as ex:
future_to_job = {ex.submit(_eval_single_job, job): job for job in batch}
for fut in as_completed(future_to_job):
done += 1
job = future_to_job[fut]
try:
result = fut.result()
results.append(result)
if output_path:
_append_result(result, output_path)
print(
f" [{done:>3}/{total}] [{mode}/{result['strategy']}] "
f"faith={result['faithfulness']:.2f} rel={result['relevancy']:.2f} "
f"ret={result['retrieval_time']:.1f}s gen={result['generation_time']:.1f}s"
)
except Exception as e:
print(f" [{done:>3}/{total}] [{mode}/{job['strategy']}] ERROR: {e}")
# Persist cache after each batch (not per-result to avoid lock contention)
if cache is not None:
_save_cache(cache)
if not is_last:
elapsed = time.perf_counter() - t_batch
wait = max(0.0, BATCH_WINDOW_S - elapsed)
if wait > 0:
print(f" [rate-limit] sleeping {wait:.1f}s before next batch...")
time.sleep(wait)
return results
def print_detailed_examples(results, n=3):
"""Print detailed output for n example evaluations."""
print(f"\n{'='*80}")
print(f"DETAILED EXAMPLES (first {n})")
print("=" * 80)
for r in results[:n]:
print(f"\nQuery: {r['query']}")
print(f"Strategy: {r['strategy']} | Mode: {r['mode']}")
print(f"Answer: {r['answer'][:300]}...")
print(f"Faithfulness: {r['faithfulness']:.3f}")
if r["faith_details"]:
for fd in r["faith_details"][:5]:
status = "+" if fd.get("supported") else "-"
print(f" {status} {fd.get('claim', '')[:100]}")
print(f"Relevancy: {r['relevancy']:.3f}")
if r["rel_alternates"]:
for alt in r["rel_alternates"]:
print(f" ~ {alt}")
print("-" * 80)
def summarize(results):
"""Group results by strategy+mode and compute mean scores including timing (§3D)."""
df = pd.DataFrame(results)
cols = [
"avg_score", "num_unique_reports",
"faithfulness", "relevancy",
"retrieval_time", "generation_time", "total_time",
]
existing = [c for c in cols if c in df.columns]
summary = df.groupby(["mode", "strategy"])[existing].mean().round(3)
return summary
def run_manual_compare_questions(
qa_items,
strategies,
jina_model,
index,
bm25_cache,
reranker,
top_k=10,
modes=("semantic", "hybrid"),
output_path=None,
use_hyde=False,
):
"""Run fixed manual QA set and print RAG answer vs reference answer."""
rows = []
for mode in modes:
for strategy in strategies:
if mode == "hybrid" and strategy not in bm25_cache:
print(f"Skipping manual compare for [hybrid/{strategy}] because BM25 index is unavailable.")
continue
print(f"\n{'='*80}")
print(f"MANUAL COMPARE | mode={mode} | strategy={strategy}")
print("=" * 80)
for item in qa_items:
qid = item["id"]
query = item["question"]
reference_answer = item["reference_answer"]
try:
if mode == "hybrid":
bm25, chunks = bm25_cache[strategy]
matches = hybrid_retrieve(
query,
strategy,
top_k=top_k,
model=jina_model,
index=index,
bm25=bm25,
chunks=chunks,
reranker=reranker,
use_hyde=use_hyde,
)
else:
matches = retrieve(query, strategy, top_k=top_k, model=jina_model, index=index)
except Exception as e:
matches = []
print(f"{qid} retrieval failed [{mode}/{strategy}]: {e}")
try:
rag_answer = generate_answer(query, matches)
except Exception as e:
rag_answer = f"Generation failed: {e}"
print(f"\n{qid}: {query}")
print("RAG_ANSWER:")
print(rag_answer)
print("REFERENCE_ANSWER:")
print(reference_answer)
print("-" * 80)
rows.append(
{
"question_id": qid,
"question": query,
"mode": mode,
"strategy": strategy,
"rag_answer": rag_answer,
"reference_answer": reference_answer,
"num_chunks": len(matches),
}
)
if output_path:
os.makedirs(os.path.dirname(output_path), exist_ok=True)
pd.DataFrame(rows).to_csv(output_path, index=False)
print(f"\nManual comparison results saved to {output_path}")
return rows
def main():
import argparse
parser = argparse.ArgumentParser(description="Run RAG evaluation")
parser.add_argument("--fresh", action="store_true",
help="Delete existing results and start from scratch")
parser.add_argument("--max-queries", type=int, default=0,
help="Limit number of evaluation queries (0 = all)")
parser.add_argument("--top-k", type=int, default=15,
help="Number of chunks to retrieve per query")
parser.add_argument("--skip-faithfulness", action="store_true",
help="Skip faithfulness scoring for faster iteration")
parser.add_argument("--skip-relevancy", action="store_true",
help="Skip relevancy scoring for faster iteration")
parser.add_argument("--no-cache", action="store_true",
help="Disable prompt cache for evaluation LLM calls")
parser.add_argument("--fast", action="store_true",
help="Quick iteration mode: fewer queries/top-k and cache enabled")
parser.add_argument("--eval-llm-provider", choices=["nvidia", "hf"], default=None,
help="Override evaluation LLM provider")
parser.add_argument("--eval-hf-model", type=str, default=None,
help="HF model ID for eval judge when provider=hf")
parser.add_argument("--use-hyde", action="store_true",
help="Enable HyDE in hybrid retrieval (better recall, slower)")
parser.add_argument("--allow-bm25-fallback", action="store_true",
help="If Pinecone retrieval fails, fall back to BM25 so evaluation can continue")
parser.add_argument("--manual-qa", action="store_true",
help="Run fixed 10-question manual comparison set")
parser.add_argument("--manual-qa-only", action="store_true",
help="Run only manual QA compare and skip standard evaluation")
parser.add_argument("--manual-qa-modes", choices=["semantic", "hybrid", "both"], default="both",
help="Retrieval modes for manual QA compare")
parser.add_argument("--manual-qa-top-k", type=int, default=10,
help="Top-k for manual QA compare")
parser.add_argument("--manual-qa-output", type=str,
default=os.path.join(BASE_DIR, "data", "manual_compare_results.csv"),
help="CSV path for manual QA compare output")
args = parser.parse_args()
if args.eval_llm_provider:
os.environ["EVAL_LLM_PROVIDER"] = args.eval_llm_provider
if args.eval_hf_model:
os.environ["EVAL_HF_MODEL"] = args.eval_hf_model
output_path = os.path.join(BASE_DIR, "data", "evaluation_results.csv")
if args.fresh and os.path.exists(output_path):
os.remove(output_path)
print("Removed existing results. Starting fresh.\n")
elif os.path.exists(output_path):
print(f"Resuming from {output_path} (use --fresh to start over)\n")
jina_model = load_model()
index = init_pinecone()
reranker = load_reranker()
eval_queries = EVAL_QUERIES
top_k = args.top_k
if args.fast:
eval_queries = EVAL_QUERIES[:5]
top_k = min(top_k, 3)
if args.max_queries and args.max_queries > 0:
eval_queries = eval_queries[:args.max_queries]
cache = None if args.no_cache else _load_cache()
strategies = available_strategies()
# Pre-build BM25 indices
bm25_cache = {}
for s in strategies:
print(f"Building BM25 index for {s}...")
try:
bm25_cache[s] = build_bm25_index(s)
except FileNotFoundError as e:
print(f" Skipping BM25 for {s}: {e}")
semantic_strategies = list(strategies)
hybrid_strategies = [s for s in strategies if s in bm25_cache]
if not hybrid_strategies:
print("Warning: No strategies have BM25 artifacts; hybrid evaluation will be skipped.")
if args.manual_qa:
manual_modes = ["semantic", "hybrid"] if args.manual_qa_modes == "both" else [args.manual_qa_modes]
run_manual_compare_questions(
qa_items=MANUAL_COMPARE_QA,
strategies=strategies,
jina_model=jina_model,
index=index,
bm25_cache=bm25_cache,
reranker=reranker,
top_k=args.manual_qa_top_k,
modes=manual_modes,
output_path=args.manual_qa_output,
use_hyde=args.use_hyde,
)
if args.manual_qa_only:
print("\nManual QA compare complete.")
return
print("\n--- Semantic-only evaluation ---")
sem_results = run_evaluation(
eval_queries, semantic_strategies, jina_model, index, mode="semantic",
bm25_cache=bm25_cache,
output_path=output_path,
top_k=top_k,
compute_faith=not args.skip_faithfulness,
compute_rel=not args.skip_relevancy,
cache=cache,
allow_bm25_fallback=args.allow_bm25_fallback or args.fast,
)
print("\n--- Hybrid evaluation ---")
if hybrid_strategies:
hyb_results = run_evaluation(
eval_queries, hybrid_strategies, jina_model, index, mode="hybrid",
bm25_cache=bm25_cache, reranker=reranker,
output_path=output_path,
top_k=top_k,
compute_faith=not args.skip_faithfulness,
compute_rel=not args.skip_relevancy,
cache=cache,
use_hyde=args.use_hyde,
allow_bm25_fallback=args.allow_bm25_fallback or args.fast,
)
else:
hyb_results = []
# Print summary from the full CSV (includes resumed + new results)
if os.path.exists(output_path):
df = pd.read_csv(output_path, on_bad_lines='skip')
summary = summarize(df.to_dict("records"))
print(f"\n{'='*80}")
print("EVALUATION SUMMARY")
print("=" * 80)
print(summary.to_string())
# Detailed examples from this run
all_results = sem_results + hyb_results
if all_results:
print_detailed_examples(all_results, n=3)
print(f"\nResults saved to {output_path}")
if __name__ == "__main__":
main()