What causes "get_embedding failed" in FoundationAgents/MetaGPT?

Enable info logs (or add logging of exp) to see the actual per-attempt exception before the final raise.

FoundationAgents/MetaGPT · error · ValueError

get_embedding failed

Error message

get_embedding failed

What it means

In the Stanford Town module, get_embedding calls OpenAI embeddings and retries 3 times (5s apart) on any exception; if all retries fail and no embedding was ever produced, it raises ValueError('get_embedding failed'). The per-attempt exceptions are only logged at info level, so the raise itself hides the root cause (bad API key, wrong model, network).

Solutions

Enable info logs (or add logging of exp) to see the actual per-attempt exception before the final raise.
Verify config.llm.api_key is set and valid, and that the embedding model name is one your key can access.
Check network/proxy connectivity to the OpenAI embeddings endpoint.
For rate limits, slow down simulation or increase retry backoff.

Example fix

// before
embedding = OpenAI(api_key=config.llm.api_key).embeddings.create(input=[text], model=model).data[0].embedding

// after (surface root cause)
client = OpenAI(api_key=config.llm.api_key)
try:
    embedding = client.embeddings.create(input=[text], model=model).data[0].embedding
except Exception as exp:
    logger.error(f"get_embedding failed permanently: {exp}")
    raise

Defensive patterns

Strategy: retry

Validate before calling

from metagpt.config2 import config

def embedding_ready(model: str) -> bool:
    return bool(config.llm.api_key)  # cheap precondition; full check needs a live call

Try / catch

from metagpt.ext.stanford_town.utils.utils import get_embedding

async def safe_embedding(text: str, model: str, attempts: int = 3):
    last = None
    for _ in range(attempts):
        try:
            return get_embedding(text, model=model)
        except ValueError as e:
            if "get_embedding failed" in str(e):
                last = e
                continue
            raise
    raise RuntimeError("embeddings unavailable; check api key/model/network") from last

Prevention

Fail fast at startup on missing api_key instead of after 3 silent retries.
Turn logger level to INFO during bring-up to capture the per-attempt exception.

When it happens

Trigger: Invalid/missing OpenAI API key in config.llm.api_key; embedding model not available to the account; network/proxy blocking api.openai.com; rate limits persisting across all 3 retries.

Common situations: Running stanford_town examples without OPENAI_API_KEY configured; using a key without access to the requested embedding model; corporate proxies or offline environments; exceeding rate limits during large simulations.

Related errors

AI-assisted analysis of FoundationAgents/MetaGPT@11cdf466d0 (2026-08-14). Data as JSON: /api/errors/3136b9f7a98472c7. Report an issue: GitHub.

Appendix: source

Thrown at metagpt/ext/stanford_town/utils/utils.py:64

                analysis_list += [row]
        return analysis_list[0], analysis_list[1:]


def get_embedding(text, model: str = "text-embedding-ada-002"):
    text = text.replace("\n", " ")
    embedding = None
    if not text:
        text = "this is blank"
    for idx in range(3):
        try:
            embedding = (
                OpenAI(api_key=config.llm.api_key).embeddings.create(input=[text], model=model).data[0].embedding
            )
        except Exception as exp:
            logger.info(f"get_embedding failed, exp: {exp}, will retry.")
            time.sleep(5)
    if not embedding:
        raise ValueError("get_embedding failed")
    return embedding


def extract_first_json_dict(data_str: str) -> Union[None, dict]:
    # Find the first occurrence of a JSON object within the string
    start_idx = data_str.find("{")
    end_idx = data_str.find("}", start_idx) + 1

    # Check if both start and end indices were found
    if start_idx == -1 or end_idx == 0:
        return None

    # Extract the first JSON dictionary
    json_str = data_str[start_idx:end_idx]

    try:
        # Attempt to parse the JSON data
        json_dict = json.loads(json_str)

View on GitHub (pinned to 11cdf466d0)