Stephen Mudere commited on
Commit
7d8e873
·
1 Parent(s): 1e816e8
Files changed (3) hide show
  1. README.md +9 -1
  2. app.py +125 -18
  3. requirements.txt +1 -0
README.md CHANGED
@@ -18,6 +18,14 @@ What changed:
18
  - routed workflow instead of the placeholder fixed-answer template
19
  - deterministic solvers for rule-based questions
20
  - targeted web retrieval for selected research questions
21
- - no paid API dependency
 
 
 
 
 
 
 
 
22
 
23
  Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
18
  - routed workflow instead of the placeholder fixed-answer template
19
  - deterministic solvers for rule-based questions
20
  - targeted web retrieval for selected research questions
21
+ - Hugging Face Inference fallback for harder tasks
22
+
23
+ Recommended Space secret:
24
+ - `HF_TOKEN`
25
+
26
+ Optional model overrides:
27
+ - `HF_CHAT_MODEL`
28
+ - `HF_ASR_MODEL`
29
+ - `HF_IMAGE_MODEL`
30
 
31
  Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
app.py CHANGED
@@ -10,6 +10,7 @@ import pandas as pd
10
  import requests
11
  from bs4 import BeautifulSoup
12
  from duckduckgo_search import DDGS
 
13
  from langgraph.graph import END, StateGraph
14
 
15
  try:
@@ -70,8 +71,13 @@ class CourseAgent:
70
  self.api_url = api_url.rstrip("/")
71
  self.session = requests.Session()
72
  self.session.headers.update(WEB_HEADERS)
 
73
  self.graph = self._build_graph()
74
 
 
 
 
 
75
  def _build_graph(self):
76
  graph = StateGraph(AgentState)
77
  graph.add_node("classify", self._classify)
@@ -204,6 +210,9 @@ class CourseAgent:
204
  if not answer and route in {"generic_web", "video_question"}:
205
  answer = self._solve_generic_web_question(question, state.get("search_results", []))
206
 
 
 
 
207
  return {"answer": answer, "notes": notes}
208
 
209
  def _finalize(self, state: AgentState) -> AgentState:
@@ -382,13 +391,14 @@ class CourseAgent:
382
 
383
  def _solve_ray_magda(self) -> str:
384
  search_results = self._search_web("Polish Everybody Loves Raymond Ray actor Magda M.", max_results=5)
385
- text = " ".join(
386
- f"{item.get('title', '')} {item.get('body', '')}" for item in search_results
 
 
 
 
 
387
  )
388
- match = re.search(r"magda m\.\s*[-: ]+\s*([A-Z][a-z]+)", text, re.I)
389
- if match:
390
- return match.group(1)
391
- return ""
392
 
393
  def _solve_1928_olympics(self) -> str:
394
  tables = pd.read_html("https://en.wikipedia.org/wiki/1928_Summer_Olympics")
@@ -456,27 +466,123 @@ class CourseAgent:
456
  match = re.search(r"isn't that hot\?\s*([^.!?]+[.!?]?)", text, re.I)
457
  if match:
458
  return match.group(1).strip().strip('"')
459
- return ""
 
 
 
 
460
 
461
  def _solve_audio_question(self, question: str, file_path: str) -> str:
462
- return ""
 
 
 
 
 
 
 
 
 
 
 
463
 
464
  def _solve_image_question(self, question: str, file_path: str) -> str:
465
- return ""
 
 
 
 
 
 
 
 
 
 
 
466
 
467
  def _solve_generic_web_question(self, question: str, search_results: list[dict[str, str]]) -> str:
468
  if not search_results:
469
  return ""
470
- best = search_results[0]
471
- snippet = best.get("body", "").strip()
472
- if not snippet:
473
- href = best.get("href", "")
 
 
 
 
 
 
 
 
 
 
474
  if href:
475
  try:
476
- snippet = self._fetch_text(href)[:200]
477
  except Exception:
478
- snippet = ""
479
- return snippet
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
480
 
481
  def _extract_video_id(self, text: str) -> str:
482
  match = re.search(r"https?://www\.youtube\.com/watch\?v=([A-Za-z0-9_-]+)", text)
@@ -569,9 +675,10 @@ with gr.Blocks() as demo:
569
  It uses:
570
  - deterministic solvers for tables, reversed text, botany classification, code output, and spreadsheets
571
  - targeted web lookups for selected Wikipedia and sports-style questions
572
- - retrieval-only fallback with no paid model dependency
573
 
574
- This version does not require any paid API key. Unsupported audio and image tasks return blank answers.
 
575
  """
576
  )
577
 
 
10
  import requests
11
  from bs4 import BeautifulSoup
12
  from duckduckgo_search import DDGS
13
+ from huggingface_hub import InferenceClient
14
  from langgraph.graph import END, StateGraph
15
 
16
  try:
 
71
  self.api_url = api_url.rstrip("/")
72
  self.session = requests.Session()
73
  self.session.headers.update(WEB_HEADERS)
74
+ self.hf_client = self._build_hf_client()
75
  self.graph = self._build_graph()
76
 
77
+ def _build_hf_client(self):
78
+ api_key = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACEHUB_API_TOKEN")
79
+ return InferenceClient(api_key=api_key)
80
+
81
  def _build_graph(self):
82
  graph = StateGraph(AgentState)
83
  graph.add_node("classify", self._classify)
 
210
  if not answer and route in {"generic_web", "video_question"}:
211
  answer = self._solve_generic_web_question(question, state.get("search_results", []))
212
 
213
+ if not answer and route not in {"generic_web", "video_question"}:
214
+ answer = self._solve_generic_web_question(question, state.get("search_results", []))
215
+
216
  return {"answer": answer, "notes": notes}
217
 
218
  def _finalize(self, state: AgentState) -> AgentState:
 
391
 
392
  def _solve_ray_magda(self) -> str:
393
  search_results = self._search_web("Polish Everybody Loves Raymond Ray actor Magda M.", max_results=5)
394
+ return self._reason_from_search_results(
395
+ question=(
396
+ "Who did the actor who played Ray in the Polish-language version of Everybody Loves Raymond "
397
+ "play in Magda M.? Give only the first name."
398
+ ),
399
+ search_results=search_results,
400
+ max_words=8,
401
  )
 
 
 
 
402
 
403
  def _solve_1928_olympics(self) -> str:
404
  tables = pd.read_html("https://en.wikipedia.org/wiki/1928_Summer_Olympics")
 
466
  match = re.search(r"isn't that hot\?\s*([^.!?]+[.!?]?)", text, re.I)
467
  if match:
468
  return match.group(1).strip().strip('"')
469
+ return self._hf_chat(
470
+ system_prompt="Extract only the direct reply to the quoted question from the transcript.",
471
+ user_prompt=f"Question: {question}\nTranscript:\n{text}",
472
+ max_tokens=60,
473
+ )
474
 
475
  def _solve_audio_question(self, question: str, file_path: str) -> str:
476
+ if not file_path:
477
+ return ""
478
+ transcript = self._transcribe_audio(file_path)
479
+ if not transcript:
480
+ return ""
481
+ return self._hf_chat(
482
+ system_prompt=(
483
+ "Answer using only the transcript. Return only the final answer text in the format requested."
484
+ ),
485
+ user_prompt=f"Question: {question}\nTranscript:\n{transcript}",
486
+ max_tokens=120,
487
+ )
488
 
489
  def _solve_image_question(self, question: str, file_path: str) -> str:
490
+ if not file_path:
491
+ return ""
492
+ image_text = self._caption_image(file_path)
493
+ if not image_text:
494
+ return ""
495
+ return self._hf_chat(
496
+ system_prompt=(
497
+ "Answer the question using the image description. Return only the final answer."
498
+ ),
499
+ user_prompt=f"Question: {question}\nImage description:\n{image_text}",
500
+ max_tokens=80,
501
+ )
502
 
503
  def _solve_generic_web_question(self, question: str, search_results: list[dict[str, str]]) -> str:
504
  if not search_results:
505
  return ""
506
+ return self._reason_from_search_results(question, search_results, max_words=40)
507
+
508
+ def _reason_from_search_results(
509
+ self,
510
+ question: str,
511
+ search_results: list[dict[str, str]],
512
+ max_words: int = 40,
513
+ ) -> str:
514
+ snippets = []
515
+ for item in search_results[:5]:
516
+ href = item.get("href", "")
517
+ body = item.get("body", "")
518
+ title = item.get("title", "")
519
+ page_text = ""
520
  if href:
521
  try:
522
+ page_text = self._fetch_text(href)[:2500]
523
  except Exception:
524
+ page_text = ""
525
+ snippets.append(f"TITLE: {title}\nURL: {href}\nSNIPPET: {body}\nPAGE: {page_text}")
526
+
527
+ return self._hf_chat(
528
+ system_prompt=(
529
+ "Answer the question using only the provided evidence. "
530
+ "Return only the exact final answer with no explanation."
531
+ ),
532
+ user_prompt=(
533
+ f"Question: {question}\n"
534
+ f"Maximum answer length: {max_words} words.\n\n"
535
+ "Evidence:\n" + "\n\n".join(snippets)
536
+ ),
537
+ max_tokens=120,
538
+ )
539
+
540
+ def _hf_chat(self, system_prompt: str, user_prompt: str, max_tokens: int = 120) -> str:
541
+ try:
542
+ output = self.hf_client.chat_completion(
543
+ model=os.getenv("HF_CHAT_MODEL", "Qwen/Qwen2.5-7B-Instruct"),
544
+ messages=[
545
+ {"role": "system", "content": system_prompt},
546
+ {"role": "user", "content": user_prompt},
547
+ ],
548
+ max_tokens=max_tokens,
549
+ temperature=0.1,
550
+ )
551
+ return output.choices[0].message.content.strip()
552
+ except Exception:
553
+ return ""
554
+
555
+ def _transcribe_audio(self, file_path: str) -> str:
556
+ try:
557
+ output = self.hf_client.automatic_speech_recognition(
558
+ file_path,
559
+ model=os.getenv("HF_ASR_MODEL", "openai/whisper-large-v3"),
560
+ )
561
+ except Exception:
562
+ return ""
563
+
564
+ text = getattr(output, "text", None)
565
+ if isinstance(text, str):
566
+ return text.strip()
567
+ if isinstance(output, str):
568
+ return output.strip()
569
+ return ""
570
+
571
+ def _caption_image(self, file_path: str) -> str:
572
+ try:
573
+ output = self.hf_client.image_to_text(
574
+ file_path,
575
+ model=os.getenv("HF_IMAGE_MODEL", "Salesforce/blip-image-captioning-large"),
576
+ )
577
+ except Exception:
578
+ return ""
579
+
580
+ text = getattr(output, "generated_text", None)
581
+ if isinstance(text, str):
582
+ return text.strip()
583
+ if isinstance(output, str):
584
+ return output.strip()
585
+ return ""
586
 
587
  def _extract_video_id(self, text: str) -> str:
588
  match = re.search(r"https?://www\.youtube\.com/watch\?v=([A-Za-z0-9_-]+)", text)
 
675
  It uses:
676
  - deterministic solvers for tables, reversed text, botany classification, code output, and spreadsheets
677
  - targeted web lookups for selected Wikipedia and sports-style questions
678
+ - Hugging Face Inference fallback for hard text, audio, and image tasks
679
 
680
+ Recommended Space secret: `HF_TOKEN`
681
+ Optional model overrides: `HF_CHAT_MODEL`, `HF_ASR_MODEL`, `HF_IMAGE_MODEL`
682
  """
683
  )
684
 
requirements.txt CHANGED
@@ -1,6 +1,7 @@
1
  beautifulsoup4
2
  duckduckgo-search
3
  gradio
 
4
  langgraph
5
  lxml
6
  openpyxl
 
1
  beautifulsoup4
2
  duckduckgo-search
3
  gradio
4
+ huggingface_hub
5
  langgraph
6
  lxml
7
  openpyxl