Panniantong/Agent-Reach · warning · ValueError

polish response exceeds 32 MiB limit

Error message

polish response exceeds 32 MiB limit

What it means

In scripts/transcribe_xiaoyuzhou.sh's embedded Python polish step: the Groq chat-completion response body exceeded 32 MiB (the script deliberately reads limit+1 bytes to detect overflow). A polish call should return a few KB of improved text; a 32 MiB body means something is pathologically wrong with the response, so it refuses to parse it.

Source

Thrown at agent_reach/scripts/transcribe_xiaoyuzhou.sh:326

    body = json.dumps({
        "model": MODEL,
        "temperature": 0.2,
        "max_completion_tokens": 8192,
        "messages": [{"role": "user", "content": PROMPT_TMPL.format(text)}],
    }).encode()
    req = urllib.request.Request(
        "https://api.groq.com/openai/v1/chat/completions",
        data=body,
        headers={
            "Authorization": f"Bearer {KEY}",
            "Content-Type": "application/json",
            "User-Agent": "agent-reach-xiaoyuzhou/1.0",
        },
    )
    with urllib.request.urlopen(req, timeout=180) as r:
        payload = r.read(32 * 1024 * 1024 + 1)
    if len(payload) > 32 * 1024 * 1024:
        raise ValueError("polish response exceeds 32 MiB limit")
    resp = json.loads(payload)
    return (
        resp["choices"][0]["message"]["content"].strip(),
        resp["choices"][0].get("finish_reason"),
    )

def polish(text, depth=0):
    try:
        out, fr = call_groq(text)
    except urllib.error.HTTPError as e:
        sys.stderr.write(f"polish HTTP {e.code}: {e.read().decode(errors='replace')[:200]}\n")
        return text  # fallback to raw
    except Exception as e:
        sys.stderr.write(f"polish error: {e}\n")
        return text
    if fr != "length" or depth >= MAX_DEPTH:
        return out
    # 输出被截断：从中点切两半递归处理

View on GitHub (pinned to 93ae1d18c3)

Solutions

1. Split long transcripts into sections and polish each separately (polish already supports iterative depth, but call it on smaller inputs)
2. Retry once — a runaway generation is usually transient
3. If reproducible, reduce the input text size passed to call_groq (summarize or trim before polishing)
4. Wrap the polish call in your own try/except ValueError to fall back to the raw transcript, mirroring the HTTPError fallback already in polish()

Example fix

# before (inside polish)
    try:
        out, fr = call_groq(text)
    except urllib.error.HTTPError as e:
        ...
        return text  # fallback to raw
# ValueError from the 32 MiB check is uncaught -> script dies

# after
    try:
        out, fr = call_groq(text)
    except (urllib.error.HTTPError, ValueError) as e:
        sys.stderr.write(f"polish failed ({e}); using raw transcript\n")
        return text  # fallback to raw

# inside polish()
try:
    out, fr = call_groq(text)
except (urllib.error.HTTPError, ValueError) as e:
    sys.stderr.write(f"polish failed ({e}); using raw transcript\n")
    return text  # fallback to raw

When it happens

Trigger: call_groq() posting an extremely large transcript to groq.com/openai/v1/chat/completions and the model/API returning a gigantic body (runaway repetition, degenerate output). Note polish() catches HTTPError but NOT this ValueError, so it propagates and kills the script.

Common situations: Very long episode transcripts (hours of text) sent in one polish request; a model looping/repeating tokens; upstream API misbehavior. Rare in practice — the 32 MiB cap is a defensive ceiling.

Related errors

• no provider key configured (need one of: {names})
• audio produced {len(chunks)} chunks; safety limit is {MAX_CH
• audio chunks total {total_chunk_bytes} bytes; safety limit i
• V2EX API response exceeds the 1 MiB safety limit
• {label} exceeds safety limit of {limit_mib:g} MiB

AI-assisted analysis of Panniantong/Agent-Reach@93ae1d18c3 (2026-08-14). Data as JSON: /api/errors/44d6f7f9d06be0ce. Report an issue: GitHub.