Claude/groq rate limits llama3 syv0w3

src/refract/ingest.py

@@ -73,6 +73,10 @@ def fetch_url(url: str) -> dict:
     """
     Fetch and extract article text from a URL using trafilatura.
     Returns a record dict with url, title, text, word_count, article_id.
+
+    No retries: a failed fetch (timeout, 404, dead link) is usually a stale
+    URL, not a transient blip, so retrying just burns time before the
+    per-article failure batch_eval.py already handles by moving on.
     """
     headers = {
         "User-Agent": (
@@ -81,43 +85,35 @@ def fetch_url(url: str) -> dict:
         )
     }
 
-    for attempt in range(1, MAX_RETRIES + 1):
-        try:
-            downloaded = trafilatura.fetch_url(url)
-            if not downloaded:
-                # trafilatura fetch failed — try requests with browser UA
-                resp = requests.get(url, headers=headers, timeout=20)
-                if resp.status_code == 404:
-                    raise ValueError(f"404 Not Found: {url}")
-                resp.raise_for_status()
-                downloaded = resp.text
-            if not downloaded:
-                raise ValueError(f"empty response from {url}")
-
-            text = trafilatura.extract(downloaded, include_comments=False, include_tables=False)
-            if not text:
-                raise ValueError(f"trafilatura extracted empty text from {url}")
-
-            meta = trafilatura.extract_metadata(downloaded)
-            title = meta.title if meta and meta.title else ""
-
-            article_id = _sha256(url, text)
-            record = {
-                "article_id": article_id,
-                "url": url,
-                "title": title,
-                "source": _source_from_url(url),
-                "text": text,
-                "word_count": len(text.split()),
-            }
-            _write_cache(article_id, record)
-            return record
-
-        except Exception as e:
-            logger.warning("fetch_url attempt %d failed for %s: %s", attempt, url, e)
-            if attempt == MAX_RETRIES:
-                raise
-            time.sleep(RETRY_BASE_DELAY * (2 ** (attempt - 1)))
+    downloaded = trafilatura.fetch_url(url)
+    if not downloaded:
+        # trafilatura fetch failed — try requests with browser UA
+        resp = requests.get(url, headers=headers, timeout=20)
+        if resp.status_code == 404:
+            raise ValueError(f"404 Not Found: {url}")
+        resp.raise_for_status()
+        downloaded = resp.text
+    if not downloaded:
+        raise ValueError(f"empty response from {url}")
+
+    text = trafilatura.extract(downloaded, include_comments=False, include_tables=False)
+    if not text:
+        raise ValueError(f"trafilatura extracted empty text from {url}")
+
+    meta = trafilatura.extract_metadata(downloaded)
+    title = meta.title if meta and meta.title else ""
+
+    article_id = _sha256(url, text)
+    record = {
+        "article_id": article_id,
+        "url": url,
+        "title": title,
+        "source": _source_from_url(url),
+        "text": text,
+        "word_count": len(text.split()),
+    }
+    _write_cache(article_id, record)
+    return record
 
 
 def fetch_guardian(query: str, page_size: int = 10, section: str = "") -> list[dict]:

src/refract/llm_client.py

@@ -234,6 +234,16 @@ def _bare_model(model: str) -> str:
     return model.split("/", 1)[1] if "/" in model else model
 
 
+# Groq's catalog namespaces some hosted models under the upstream provider
+# (e.g. "openai/gpt-oss-120b"), which collides with this codebase's own
+# "groq/" routing prefix used to disambiguate from Cerebras's identically
+# named "gpt-oss-120b". Map our routing alias to Groq's real model id before
+# calling the API — otherwise the bare-stripped "gpt-oss-120b" 404s.
+_GROQ_MODEL_ALIASES = {
+    "gpt-oss-120b": "openai/gpt-oss-120b",
+}
+
+
 def _provider(model: str) -> str:
     """
     Infer API provider from model name.
@@ -256,7 +266,9 @@ def _provider(model: str) -> str:
 
 
 def _call_groq(prompt: str, model: str, system: str = "", expect_json: bool = True, temperature: float = 0.0) -> tuple[dict | str, int]:
-    return _call_openai_compat(prompt, _bare_model(model), system, expect_json, temperature, GROQ_BASE, GROQ_API_KEY, "groq")
+    bare = _bare_model(model)
+    real_model = _GROQ_MODEL_ALIASES.get(bare, bare)
+    return _call_openai_compat(prompt, real_model, system, expect_json, temperature, GROQ_BASE, GROQ_API_KEY, "groq")
 
 
 def _parse_json_response(raw: str) -> Any:
@@ -331,7 +343,7 @@ def _call_openai_compat(
     messages.append({"role": "user", "content": prompt})
 
     body: dict[str, Any] = {
-        "model": _bare_model(model),
+        "model": model,
         "messages": messages,
         "temperature": temperature,
     }