mbochniak01 Claude Sonnet 4.6 commited on
Commit
27156ca
·
1 Parent(s): 14d263b

Add joke short-circuit and reset attempt handler

Browse files

Joke queries (joke, funny, pun, dad joke, etc.) get a canned domain-themed
joke instead of going through RAG — avoids faithfulness false positives on
creative/off-topic requests.

Reset/jailbreak attempts (forget, blank slate, ignore context, etc.) also
route to the joke bank: light-hearted deflection, no compliance risk.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Files changed (1) hide show
  1. backend/pipeline.py +61 -0
backend/pipeline.py CHANGED
@@ -14,6 +14,7 @@ Knowledge base formats supported:
14
  import csv
15
  import json
16
  import logging
 
17
  import re
18
  from dataclasses import dataclass, field
19
  from pathlib import Path
@@ -259,6 +260,57 @@ def _build_context(docs: list[RetrievedDoc]) -> str:
259
  return "\n\n".join(f"[{d.title}]\n{d.content.strip()}" for d in docs)
260
 
261
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
262
  # ---------------------------------------------------------------------------
263
  # Generation
264
  # ---------------------------------------------------------------------------
@@ -303,6 +355,15 @@ def run(
303
  ) -> PipelineResult:
304
  """Retrieve relevant KB docs, generate a grounded answer, and grade it."""
305
  domain = domain_for(client)
 
 
 
 
 
 
 
 
 
306
  embedder = get_embedder()
307
  index = _build_index(domain, embedder)
308
 
 
14
  import csv
15
  import json
16
  import logging
17
+ import random
18
  import re
19
  from dataclasses import dataclass, field
20
  from pathlib import Path
 
260
  return "\n\n".join(f"[{d.title}]\n{d.content.strip()}" for d in docs)
261
 
262
 
263
+ # ---------------------------------------------------------------------------
264
+ # Joke short-circuit
265
+ # ---------------------------------------------------------------------------
266
+
267
+ _JOKE_PATTERN = re.compile(
268
+ r"\b(joke|jokes|funny|laugh|humor|humour|dad joke|daddy joke|pun|make me (laugh|smile))\b",
269
+ re.IGNORECASE,
270
+ )
271
+
272
+ _DOMAIN_JOKES: dict[str, list[str]] = {
273
+ "retail": [
274
+ "Why did the inventory go to therapy? It had too many unresolved stockouts.",
275
+ "What do you call a shelf that never runs out? A shelf-fulfilling prophecy.",
276
+ "Why don't retail systems ever get lonely? Because they're always in stock-ing.",
277
+ "I tried to write a joke about planograms… but I couldn't find the right placement.",
278
+ "Why did the out-of-stock alert break up with the replenishment order? It said: 'You always come too late.'",
279
+ ],
280
+ "pharma": [
281
+ "Why did the adverse event report go to the doctor? It had too many side effects.",
282
+ "What do you call a drug that's always on time? Punctu-pill.",
283
+ "Why are pharmacovigilance teams great at parties? They always follow up.",
284
+ "I asked the formulary for a joke. It said: 'Prior authorization required.'",
285
+ "Why did the clinical trial fail to finish the joke? Insufficient sample size.",
286
+ ],
287
+ }
288
+
289
+ _FALLBACK_JOKES = [
290
+ "Why did the AI refuse to tell a joke? Insufficient context. Please provide grounding documents.",
291
+ "I'm not great at jokes — my training data was mostly compliance documents.",
292
+ "Why did the chatbot cross the road? To reduce hallucination on the other side.",
293
+ ]
294
+
295
+ _RESET_PATTERNS = re.compile(
296
+ r"\b(forget|ignore|pretend|blank slate|no context|disregard|system prompt|jailbreak)\b",
297
+ re.IGNORECASE,
298
+ )
299
+
300
+
301
+ def _is_joke_query(query: str) -> bool:
302
+ return bool(_JOKE_PATTERN.search(query))
303
+
304
+
305
+ def _joke_response(domain: str) -> str:
306
+ jokes = _DOMAIN_JOKES.get(domain, _FALLBACK_JOKES)
307
+ return random.choice(jokes)
308
+
309
+
310
+ def _is_reset_attempt(query: str) -> bool:
311
+ return bool(_RESET_PATTERNS.search(query))
312
+
313
+
314
  # ---------------------------------------------------------------------------
315
  # Generation
316
  # ---------------------------------------------------------------------------
 
355
  ) -> PipelineResult:
356
  """Retrieve relevant KB docs, generate a grounded answer, and grade it."""
357
  domain = domain_for(client)
358
+
359
+ if _is_joke_query(query) or _is_reset_attempt(query):
360
+ answer = _joke_response(domain)
361
+ report = grade(query=query, response=answer, context=answer, client=client)
362
+ return PipelineResult(
363
+ query=query, client=client, answer=answer,
364
+ retrieved_docs=[], grade_report=report, context_used="",
365
+ )
366
+
367
  embedder = get_embedder()
368
  index = _build_index(domain, embedder)
369