feat: enhance answer extraction and normalization logic, improve error handling in LLM invocation, and optimize YouTube transcript retrieval

agent.py

@@ -46,21 +46,20 @@ SYSTEM_PROMPT = """RESEARCH STRATEGY:
 TOOL SELECTION:
 - web_search: General web search for facts, URLs, news. Returns 5 results with title+URL.
 - wiki_search: Find Wikipedia pages on a topic (returns 2 short snippets only).
-- wiki_page: Read a FULL Wikipedia article by exact title. Use AFTER wiki_search when you need complete content.
+- wiki_page: Read a FULL Wikipedia article by exact title. Use AFTER wiki_search.
 - browse_url: Read any webpage's full text content from a URL.
 - read_file: Text files (.txt, .py, .md), PDFs. NOT for spreadsheets, audio, or images.
 - parse_spreadsheet: Excel (.xlsx) and CSV files. Always use this for spreadsheets.
-- transcribe_audio: Audio files (.mp3, .wav). The ONLY tool for audio — use it immediately.
-- python_repl: Calculations, data analysis, PDF parsing, file processing. Variables persist between calls. Use libraries already installed (pandas, numpy, PIL, PyPDF2).
+- transcribe_audio: Audio files (.mp3, .wav). Call once, use the returned text immediately.
+- python_repl: Calculations, data analysis, file processing. Variables persist between calls.
 - get_youtube_transcript: YouTube video transcripts only.
 
-RULES:
+EFFICIENCY RULES (save tool calls):
+- After transcribe_audio returns text, USE IT. Do NOT call read_file on audio, do NOT search for audio files with python_repl.
+- After read_file shows a .py script, run it with exec(open(path).read()) — do NOT rewrite the code.
+- After getting a tool result, analyze it. Do not search again with slightly different queries.
 - Never call the same tool with the same arguments twice.
-- When wiki_search returns short snippets, follow up with wiki_page for full content.
-- After getting tool results, analyze them — don't immediately search again with slightly different queries.
-- Audio files can ONLY be read with transcribe_audio. Do not use read_file on .mp3/.wav.
-- Images (.png/.jpg) cannot be analyzed — no vision model. Skip file-based analysis.
-- For spreadsheets: parse_spreadsheet to view, then python_repl to compute if needed.
+- Images (.png/.jpg) cannot be analyzed — no vision available. Skip them completely.
 - Don't install packages in python_repl — use what's already available."""
 
 

app.py

@@ -20,13 +20,20 @@ def extract_answer(content) -> str:
         cleaned = content.strip()
         if not cleaned:
             return ""
-        match = re.search(r'FINAL ANSWER:\s*(.+?)(?:\n|$)', cleaned, re.IGNORECASE)
+        cleaned = re.sub(r'```[\w]*\n?', '', cleaned)
+        match = re.search(r'^FINAL ANSWER:\s*(.+?)(?:\n\n|\Z|(?=\nFINAL))', cleaned, re.IGNORECASE | re.MULTILINE | re.DOTALL)
         if match:
             return match.group(1).strip()
-        match = re.search(r'Answer:\s*(.+?)(?:\n|$)', cleaned, re.IGNORECASE)
+        match = re.search(r'FINAL ANSWER:\s*(.+?)(?:\n\n|\Z|(?=\nFINAL))', cleaned, re.IGNORECASE | re.DOTALL)
         if match:
             return match.group(1).strip()
-        match = re.search(r'(?:the\s+)?answer\s+is\s*:?\s*(.+?)(?:\.|$)', cleaned, re.IGNORECASE)
+        match = re.search(r'Answer:\s*(.+?)(?:\n\n|\Z)', cleaned, re.IGNORECASE | re.DOTALL)
+        if match:
+            return match.group(1).strip()
+        match = re.search(r'(?:the\s+)?answer\s+is\s*:?\s*(.+?)(?:\n\n|\Z)', cleaned, re.IGNORECASE | re.DOTALL)
+        if match:
+            return match.group(1).strip()
+        match = re.search(r'(?:^|\n)(?:Result|Output|Therefore)[:\s]+(.+)', cleaned, re.IGNORECASE)
         if match:
             return match.group(1).strip()
         lines = [l.strip() for l in cleaned.split('\n') if l.strip()]
@@ -38,8 +45,11 @@ def extract_answer(content) -> str:
 
 def normalize_answer(s: str) -> str:
     s = s.strip().lower()
-    s = re.sub(r'^[\$€£¥]', '', s)
-    s = s.replace(',', '')
+    s = re.sub(r'^[\$€£¥]+', '', s)
+    s = re.sub(r'[.,!?;:]+$', '', s)
+    s = s.replace(',', ' ')
+    s = re.sub(r'\s+', ' ', s)
+    s = re.sub(r'^(a |an |the )', '', s)
     s = s.rstrip('*')
     s = s.strip()
     return s
@@ -169,7 +179,7 @@ def run_and_submit_all(profile: Optional[gr.OAuthProfile] = None):
                     ".png": "Tip: This is an image — no vision model available. Skip file-based analysis.",
                     ".jpg": "Tip: This is an image — no vision model available. Skip file-based analysis.",
                     ".jpeg": "Tip: This is an image — no vision model available. Skip file-based analysis.",
-                    ".py": "Tip: This is a Python script. Use read_file to view, then python_repl to run it.",
+                    ".py": "Tip: This is a Python script. Use read_file to view, then python_repl with exec(open(path).read()) to run it directly.",
                     ".pdf": "Tip: This is a PDF. Use read_file or python_repl with PyPDF2 to read it.",
                 }
                 _hint = _tips.get(_ext, "")

gaia_results.csv

@@ -14,11 +14,11 @@ cca530fc-4052-43b2-b130-b30968d8aa44,Review the chess position provided in the i
 |d|b|e|b|e|d|
 |e|d|b|a|d|c|
 
-provide the subset of S involved in any possible counter-examples that prove * is not commutative. Provide your answer as a comma separated list of the elements in the set in alphabetical order.","b, e","b, e",True
+provide the subset of S involved in any possible counter-examples that prove * is not commutative. Provide your answer as a comma separated list of the elements in the set in alphabetical order.","b,e","b, e",False
 9d191bce-651d-4746-be2d-7ef8ecadb9c2,"Examine the video at https://www.youtube.com/watch?v=1htKBjuUWec.
 
-What does Teal'c say in response to the question ""Isn't that hot?""","Teal'c responds with **""Extremely.""**",Extremely,False
-cabe07ed-9eca-40ea-8ead-410ef5e83f91,What is the surname of the equine veterinarian mentioned in 1.E Exercises from the chemistry materials licensed by Marisa Alviar-Agnew & Henry Agnew under the CK-12 license in LibreText's Introductory Chemistry materials as compiled 08/21/2023?,Louvrier,Louvrier,True
+What does Teal'c say in response to the question ""Isn't that hot?""","Teal'c responds with **""Indeed.""**",Extremely,False
+cabe07ed-9eca-40ea-8ead-410ef5e83f91,What is the surname of the equine veterinarian mentioned in 1.E Exercises from the chemistry materials licensed by Marisa Alviar-Agnew & Henry Agnew under the CK-12 license in LibreText's Introductory Chemistry materials as compiled 08/21/2023?,,Louvrier,False
 3cef3a44-215e-4aed-8e3b-b1e3f08063b7,"I'm making a grocery list for my mom, but she's a professor of botany and she's a real stickler when it comes to categorizing things. I need to add different foods to different categories on the grocery list, but if I make a mistake, she won't buy anything inserted in the wrong category. Here's the list I have so far:
 
 milk, eggs, flour, whole bean coffee, Oreos, sweet potatoes, fresh basil, plums, green beans, rice, corn, bell pepper, whole allspice, acorns, broccoli, celery, zucchini, lettuce, peanuts
@@ -30,14 +30,14 @@ In your response, please only list the ingredients, not any measurements. So if
 
 Please format your response as a comma separated list of ingredients. Also, please alphabetize the ingredients.",,"cornstarch, freshly squeezed lemon juice, granulated sugar, pure vanilla extract, ripe strawberries",False
 305ac316-eef6-4446-960a-92d80d542f82,Who did the actor who played Ray in the Polish-language version of Everybody Loves Raymond play in Magda M.? Give only the first name.,,Wojciech,False
-f918266a-b3e0-4914-865d-4faa564f1aef,What is the final numeric output from the attached Python code?,None,0,False
+f918266a-b3e0-4914-865d-4faa564f1aef,What is the final numeric output from the attached Python code?,,0,False
 3f57289b-8c60-48be-bd80-01f8099ca449,How many at bats did the Yankee with the most walks in the 1977 regular season have that same season?,,519,False
 1f975693-876d-457b-a649-393859e79bf3,"Hi, I was out sick from my classes on Friday, so I'm trying to figure out what I need to study for my Calculus mid-term next week. My friend from class sent me an audio recording of Professor Willowbrook giving out the recommended reading for the test, but my headphones are broken :(
 
 Could you please listen to the recording for me and tell me the page numbers I'm supposed to go over? I've attached a file called Homework.mp3 that has the recording. Please provide just the page numbers as a comma-delimited list. And please provide the list in ascending order.",,"132, 133, 134, 197, 245",False
-840bfca7-4f7b-481a-8794-c560c340185d,"On June 6, 2023, an article by Carolyn Collins Petersen was published in Universe Today. This article mentions a team that produced a paper about their observations, linked at the bottom of the article. Find this paper. Under what NASA award number was the work performed by R. G. Arendt supported by?",The work performed by R. G. Arendt was supported by NASA award number **80GSFC21M0002**.,80GSFC21M0002,False
-bda648d7-d618-4883-88f4-3466eabd860e,Where were the Vietnamese specimens described by Kuznetzov in Nedoshivina's 2010 paper eventually deposited? Just give me the city name without abbreviations.,Saint Petersburg,Saint Petersburg,True
+840bfca7-4f7b-481a-8794-c560c340185d,"On June 6, 2023, an article by Carolyn Collins Petersen was published in Universe Today. This article mentions a team that produced a paper about their observations, linked at the bottom of the article. Find this paper. Under what NASA award number was the work performed by R. G. Arendt supported by?",,80GSFC21M0002,False
+bda648d7-d618-4883-88f4-3466eabd860e,Where were the Vietnamese specimens described by Kuznetzov in Nedoshivina's 2010 paper eventually deposited? Just give me the city name without abbreviations.,,Saint Petersburg,False
 cf106601-ab4f-4af9-b045-5295fe67b37d,"What country had the least number of athletes at the 1928 Summer Olympics? If there's a tie for a number of athletes, return the first in alphabetical order. Give the IOC country code as your answer.",CUB,CUB,True
-a0c07678-e491-4bbc-8f0b-07405144218f,"Who are the pitchers with the number before and after Taishō Tamai's number as of July 2023? Give them to me in the form Pitcher Before, Pitcher After, use their last names only, in Roman characters.",,"Yoshida, Uehara",False
+a0c07678-e491-4bbc-8f0b-07405144218f,"Who are the pitchers with the number before and after Taishō Tamai's number as of July 2023? Give them to me in the form Pitcher Before, Pitcher After, use their last names only, in Roman characters.",ERROR: 'charmap' codec can't encode character '\u014d' in position 55: character maps to <undefined>,"Yoshida, Uehara",False
 7bd855d8-463d-4ed5-93ca-5fe35145f733,The attached Excel file contains the sales of menu items for a local fast-food chain. What were the total sales that the chain made from food (not including drinks)? Express your answer in USD with two decimal places.,89706.00,89706.00,True
 5a0c1adf-205e-4841-a666-7c3ef95def9d,What is the first name of the only Malko Competition recipient from the 20th Century (after 1977) whose nationality on record is a country that no longer exists?,,Claus,False

gaia_results.json

@@ -1,6 +1,6 @@
 {
-  "score": 35.0,
-  "correct": 7,
+  "score": 20.0,
+  "correct": 4,
   "total": 20,
   "results": [
     {
@@ -41,23 +41,23 @@
     {
       "task_id": "6f37996b-2ac7-44b0-8e68-6d28256631b4",
       "question": "Given this table defining * on the set S = {a, b, c, d, e}\n\n|*|a|b|c|d|e|\n|---|---|---|---|---|---|\n|a|a|b|c|b|d|\n|b|b|c|a|e|c|\n|c|c|a|b|b|a|\n|d|b|e|b|e|d|\n|e|d|b|a|d|c|\n\nprovide the subset of S involved in any possible counter-examples that prove * is not commutative. Provide your answer as a comma separated list of the elements in the set in alphabetical order.",
-      "submitted_answer": "b, e",
+      "submitted_answer": "b,e",
       "ground_truth": "b, e",
-      "correct": true
+      "correct": false
     },
     {
       "task_id": "9d191bce-651d-4746-be2d-7ef8ecadb9c2",
       "question": "Examine the video at https://www.youtube.com/watch?v=1htKBjuUWec.\n\nWhat does Teal'c say in response to the question \"Isn't that hot?\"",
-      "submitted_answer": "Teal'c responds with **\"Extremely.\"**",
+      "submitted_answer": "Teal'c responds with **\"Indeed.\"**",
       "ground_truth": "Extremely",
       "correct": false
     },
     {
       "task_id": "cabe07ed-9eca-40ea-8ead-410ef5e83f91",
       "question": "What is the surname of the equine veterinarian mentioned in 1.E Exercises from the chemistry materials licensed by Marisa Alviar-Agnew & Henry Agnew under the CK-12 license in LibreText's Introductory Chemistry materials as compiled 08/21/2023?",
-      "submitted_answer": "Louvrier",
+      "submitted_answer": "",
       "ground_truth": "Louvrier",
-      "correct": true
+      "correct": false
     },
     {
       "task_id": "3cef3a44-215e-4aed-8e3b-b1e3f08063b7",
@@ -83,7 +83,7 @@
     {
       "task_id": "f918266a-b3e0-4914-865d-4faa564f1aef",
       "question": "What is the final numeric output from the attached Python code?",
-      "submitted_answer": "None",
+      "submitted_answer": "",
       "ground_truth": "0",
       "correct": false
     },
@@ -104,16 +104,16 @@
     {
       "task_id": "840bfca7-4f7b-481a-8794-c560c340185d",
       "question": "On June 6, 2023, an article by Carolyn Collins Petersen was published in Universe Today. This article mentions a team that produced a paper about their observations, linked at the bottom of the article. Find this paper. Under what NASA award number was the work performed by R. G. Arendt supported by?",
-      "submitted_answer": "The work performed by R.\u202fG.\u202fArendt was supported by NASA award number **80GSFC21M0002**.",
+      "submitted_answer": "",
       "ground_truth": "80GSFC21M0002",
       "correct": false
     },
     {
       "task_id": "bda648d7-d618-4883-88f4-3466eabd860e",
       "question": "Where were the Vietnamese specimens described by Kuznetzov in Nedoshivina's 2010 paper eventually deposited? Just give me the city name without abbreviations.",
-      "submitted_answer": "Saint Petersburg",
+      "submitted_answer": "",
       "ground_truth": "Saint Petersburg",
-      "correct": true
+      "correct": false
     },
     {
       "task_id": "cf106601-ab4f-4af9-b045-5295fe67b37d",
@@ -125,7 +125,7 @@
     {
       "task_id": "a0c07678-e491-4bbc-8f0b-07405144218f",
       "question": "Who are the pitchers with the number before and after Taish\u014d Tamai's number as of July 2023? Give them to me in the form Pitcher Before, Pitcher After, use their last names only, in Roman characters.",
-      "submitted_answer": "",
+      "submitted_answer": "ERROR: 'charmap' codec can't encode character '\\u014d' in position 55: character maps to <undefined>",
       "ground_truth": "Yoshida, Uehara",
       "correct": false
     },

llm/client.py

@@ -1,5 +1,6 @@
 import os
 import time
+import random
 
 from langchain_core.messages import AIMessage
 from llm.providers import PROVIDERS
@@ -7,6 +8,25 @@ from llm.providers import PROVIDERS
 PROVIDER_ORDER = [p.strip() for p in os.getenv("LLM_PROVIDER_ORDER", "opencode_zen, groq").split(",")]
 
 
+def _is_rate_limit(e: Exception) -> bool:
+    err = str(e).lower()
+    return any(x in err for x in ("rate limit", "429", "quota", "resource ex"))
+
+
+def _is_retryable(e: Exception) -> bool:
+    err = str(e).lower()
+    if _is_rate_limit(e):
+        return True
+    return any(x in err for x in ("500", "502", "503", "504", "timeout", "connection", "temporary", "bad gateway",
+                                   "service unavailable", "internal server error", "server error", "temporarily"))
+
+
+def _is_unrecoverable(e: Exception) -> bool:
+    err = str(e).lower()
+    return any(x in err for x in ("payment required", "402", "tool_use_failed", "model_not_found",
+                                   "too large", "413", "400", "401", "403", "404"))
+
+
 def invoke_llm(messages, tools, fallback_count=0, _degraded=None):
     if _degraded is None:
         _degraded = {}
@@ -27,19 +47,23 @@ def invoke_llm(messages, tools, fallback_count=0, _degraded=None):
             print(f"Invoking {provider_name} with {model_name}", flush=True)
 
             retries = 0
-            while retries < 2:
+            max_retries = 3
+            while retries < max_retries:
                 try:
                     return provider.invoke(messages, tools, model_name)
                 except Exception as e:
                     err_str = str(e)
-                    err = err_str.lower()
-                    if any(x in err for x in ("rate limit", "429", "quota", "resource ex")):
-                        print(f"{provider_name}/{model_name} rate limited, waiting...", flush=True)
-                        time.sleep(65)
-                        retries += 1
-                    elif any(x in err for x in ("payment required", "402", "tool_use_failed", "model_not_found", "too large", "413")):
-                        print(f"{provider_name}/{model_name} skip, trying next", flush=True)
+                    if _is_unrecoverable(e):
+                        print(f"{provider_name}/{model_name} skip ({err_str[:100]})", flush=True)
                         break
+                    elif _is_retryable(e):
+                        if _is_rate_limit(e):
+                            delay = 65
+                        else:
+                            delay = min(2 ** retries + random.uniform(0, 1), 30)
+                        print(f"{provider_name}/{model_name} retryable ({err_str[:100]}), retry {retries+1}/{max_retries} after {delay:.0f}s", flush=True)
+                        time.sleep(delay)
+                        retries += 1
                     else:
                         print(f"{provider_name}/{model_name} error: {type(e).__name__}: {err_str[:150]}", flush=True)
                         break

run_local.py

@@ -19,13 +19,20 @@ def extract_answer(content) -> str:
         cleaned = content.strip()
         if not cleaned:
             return ""
-        match = re.search(r'FINAL ANSWER:\s*(.+?)(?:\n|$)', cleaned, re.IGNORECASE)
+        cleaned = re.sub(r'```[\w]*\n?', '', cleaned)
+        match = re.search(r'^FINAL ANSWER:\s*(.+?)(?:\n\n|\Z|(?=\nFINAL))', cleaned, re.IGNORECASE | re.MULTILINE | re.DOTALL)
         if match:
             return match.group(1).strip()
-        match = re.search(r'Answer:\s*(.+?)(?:\n|$)', cleaned, re.IGNORECASE)
+        match = re.search(r'FINAL ANSWER:\s*(.+?)(?:\n\n|\Z|(?=\nFINAL))', cleaned, re.IGNORECASE | re.DOTALL)
         if match:
             return match.group(1).strip()
-        match = re.search(r'(?:the\s+)?answer\s+is\s*:?\s*(.+?)(?:\.|$)', cleaned, re.IGNORECASE)
+        match = re.search(r'Answer:\s*(.+?)(?:\n\n|\Z)', cleaned, re.IGNORECASE | re.DOTALL)
+        if match:
+            return match.group(1).strip()
+        match = re.search(r'(?:the\s+)?answer\s+is\s*:?\s*(.+?)(?:\n\n|\Z)', cleaned, re.IGNORECASE | re.DOTALL)
+        if match:
+            return match.group(1).strip()
+        match = re.search(r'(?:^|\n)(?:Result|Output|Therefore)[:\s]+(.+)', cleaned, re.IGNORECASE)
         if match:
             return match.group(1).strip()
         lines = [l.strip() for l in cleaned.split('\n') if l.strip()]
@@ -37,8 +44,11 @@ def extract_answer(content) -> str:
 
 def normalize_answer(s: str) -> str:
     s = s.strip().lower()
-    s = re.sub(r'^[\$€£¥]', '', s)
-    s = s.replace(',', '')
+    s = re.sub(r'^[\$€£¥]+', '', s)
+    s = re.sub(r'[.,!?;:]+$', '', s)
+    s = s.replace(',', ' ')
+    s = re.sub(r'\s+', ' ', s)
+    s = re.sub(r'^(a |an |the )', '', s)
     s = s.rstrip('*')
     s = s.strip()
     return s
@@ -125,7 +135,7 @@ def main():
                     ".png": "Tip: This is an image — no vision model available. Skip file-based analysis.",
                     ".jpg": "Tip: This is an image — no vision model available. Skip file-based analysis.",
                     ".jpeg": "Tip: This is an image — no vision model available. Skip file-based analysis.",
-                    ".py": "Tip: This is a Python script. Use read_file to view, then python_repl to run it.",
+                    ".py": "Tip: This is a Python script. Use read_file to view, then python_repl with exec(open(path).read()) to run it directly.",
                     ".pdf": "Tip: This is a PDF. Use read_file or python_repl with PyPDF2 to read it.",
                 }
                 _hint = _tips.get(_ext, "")

tools/audio.py

@@ -1,13 +1,40 @@
+import os
+import hashlib
 from langchain_core.tools import tool
 
+_whisper_model = None
+_transcript_cache = {}
+
 
 @tool
 def transcribe_audio(path: str) -> str:
-    """Transcribe audio file to text."""
+    """Transcribe audio file to text. Results are cached per file. Call this ONCE for each audio file and use the returned text directly. Do not call read_file or search for the file afterwards."""
+    global _whisper_model
+
+    if not path or not os.path.exists(path):
+        return "ERROR: Audio file not found"
+
+    abs_path = os.path.abspath(path)
+    mtime = os.path.getmtime(abs_path)
+    cache_key = f"{abs_path}::{mtime}"
+
+    cached = _transcript_cache.get(cache_key)
+    if cached is not None:
+        return cached
+
     try:
         import whisper
-        model = whisper.load_model("base")
-        result = model.transcribe(path)
-        return result["text"][:5000] or "NO_TRANSCRIPTION"
+        import torch
+
+        if _whisper_model is None:
+            device = "cuda" if torch.cuda.is_available() else "cpu"
+            _whisper_model = whisper.load_model("base", device=device)
+
+        result = _whisper_model.transcribe(abs_path)
+        text = result["text"][:5000] or "NO_TRANSCRIPTION"
+        if result["text"] and len(result["text"]) > 5000:
+            text += "\n...[transcript truncated at 5000 characters]"
+        _transcript_cache[cache_key] = text
+        return text
     except Exception as e:
         return f"AUDIO_TRANSCRIPTION_ERROR: {e}"

tools/file/reader.py

@@ -1,4 +1,5 @@
 import os
+from pathlib import Path
 import fitz
 from langchain_community.document_loaders import UnstructuredFileLoader
 from langchain_community.document_loaders.image import UnstructuredImageLoader
@@ -40,6 +41,8 @@ def read_file(path: str) -> str:
                 loader = UnstructuredFileLoader(path)
                 docs = loader.load()
                 content = "\n\n".join([doc.page_content for doc in docs])
+        elif ext == ".py":
+            content = Path(path).read_text(encoding="utf-8", errors="replace")
         else:
             loader = UnstructuredFileLoader(path)
             docs = loader.load()

tools/python.py

@@ -9,6 +9,7 @@ def python_repl(code: str) -> str:
     The code should be a valid python script that prints the final result.
     You can use libraries like pandas, numpy, PIL, etc.
     IMPORTANT: Variables persist between calls to this tool (same Python process). You can define a variable in one call and use it in the next.
+    To run a .py file: exec(open('/path/to/file.py').read())
     Example: print(df.head()) or print(2 + 2)"""
     try:
         old_stdout = sys.stdout

tools/web/search.py

@@ -1,20 +1,27 @@
+import time
 from langchain_core.tools import tool
 
 
 @tool
 def web_search(keywords: str) -> str:
     """Search the web. Use this for finding information, facts, URLs, and general web content. Returns title, URL, and snippet for each result."""
-    try:
-        from ddgs import DDGS
-        results = list(DDGS().text(keywords, max_results=5))
-        if not results:
-            return "NO_RESULTS"
-        formatted = []
-        for r in results:
-            title = r.get("title", "")
-            url = r.get("href", "")
-            body = r.get("body", "")[:300]
-            formatted.append(f"Title: {title}\nURL: {url}\nContent: {body}")
-        return "\n\n".join(formatted)
-    except Exception as e:
-        return f"SEARCH_ERROR: {e}"
+    headers = {"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36"}
+    last_err = None
+    for attempt in range(3):
+        try:
+            from ddgs import DDGS
+            results = list(DDGS(headers=headers).text(keywords, max_results=5, backend="html"))
+            if not results:
+                return "NO_RESULTS"
+            formatted = []
+            for r in results:
+                title = r.get("title", "")
+                url = r.get("href", "")
+                body = r.get("body", "")[:300]
+                formatted.append(f"Title: {title}\nURL: {url}\nContent: {body}")
+            return "\n\n".join(formatted)
+        except Exception as e:
+            last_err = e
+            if attempt < 2:
+                time.sleep(2 ** attempt)
+    return f"SEARCH_ERROR: {last_err}"

tools/youtube.py

@@ -4,18 +4,49 @@ from pathlib import Path
 from langchain_core.tools import tool
 
 
+def _parse_vtt(content: str) -> str:
+    lines = []
+    for l in content.splitlines():
+        stripped = l.strip()
+        if not stripped:
+            continue
+        if stripped.startswith(('<', '-->', 'WEBVTT', 'NOTE', 'STYLE', '::')):
+            continue
+        if stripped.isdigit():
+            continue
+        if 'align:' in stripped or 'position:' in stripped:
+            continue
+        lines.append(stripped)
+    return "\n".join(lines)
+
+
+def _try_download(url: str, tmp: str, extra_args: list) -> list[Path]:
+    cmd = ["yt-dlp", "--skip-download", "-o", f"{tmp}/video", url] + extra_args
+    result = subprocess.run(cmd, capture_output=True, timeout=120)
+    if result.returncode != 0:
+        return []
+    return list(Path(tmp).glob("*.vtt"))
+
+
 @tool
 def get_youtube_transcript(url: str) -> str:
-    """Get YouTube transcript."""
+    """Get YouTube transcript. Tries manual captions first, then auto-generated, falling back through language variants."""
+    strategies = [
+        ["--write-subs", "--sub-langs", "en.*,en"],
+        ["--write-auto-subs", "--sub-langs", "en.*,en"],
+        ["--write-auto-subs", "--sub-langs", "a.*"],
+    ]
     try:
         with tempfile.TemporaryDirectory() as tmp:
-            cmd = ["yt-dlp", "--skip-download", "--write-auto-subs", "--sub-lang", "en", "-o", f"{tmp}/video", url]
-            subprocess.run(cmd, capture_output=True, timeout=60)
-            vtt_files = list(Path(tmp).glob("*.vtt"))
-            if vtt_files:
-                content = vtt_files[0].read_text(encoding="utf-8", errors="replace")
-                lines = [l for l in content.splitlines() if l and not l.startswith(('<', '-->', 'WEBVTT')) and not l.isdigit()]
-                return "\n".join(lines)[:15000] or "NO_TRANSCRIPT"
+            for strat in strategies:
+                vtt_files = _try_download(url, tmp, strat)
+                if vtt_files:
+                    content = vtt_files[0].read_text(encoding="utf-8", errors="replace")
+                    parsed = _parse_vtt(content)
+                    if parsed.strip():
+                        return parsed[:15000]
             return "NO_SUBTITLES"
+    except subprocess.TimeoutExpired:
+        return "TRANSCRIPT_ERROR: timeout"
     except Exception as e:
         return f"TRANSCRIPT_ERROR: {e}"