BerriAI/litellm · error · ValueError

Missing expected key in embedding response: {e}

Error message

Missing expected key in embedding response: {e}

What it means

ValueError raised while LiteLLM converts an embedding response into the cacheable per-item dict form: a KeyError occurred while accessing expected attributes/keys of the response data. The f-string interpolates the missing key name (the KeyError), telling you exactly which field was absent. It signals a provider response shape that does not match what the caching layer expects for embeddings.

Source

Thrown at litellm/caching/caching.py:726

                    "embedding": data.get("embedding"),
                    "index": data.get("index"),
                    "object": data.get("object"),
                    "model": model,
                    "prompt_tokens": prompt_tokens,
                    "prompt_tokens_details": prompt_tokens_details,
                }
            else:
                data = vars(embedding_response)
                return {
                    "embedding": data.get("embedding"),
                    "index": data.get("index"),
                    "object": data.get("object"),
                    "model": model,
                    "prompt_tokens": prompt_tokens,
                    "prompt_tokens_details": prompt_tokens_details,
                }
        except KeyError as e:
            raise ValueError(f"Missing expected key in embedding response: {e}")

    def _get_per_item_prompt_tokens_details(
        self,
        result: EmbeddingResponse,
        idx_in_result_data: int,
    ) -> dict | None:
        """
        Extract per-item prompt_tokens_details from a response for caching.

        For single-item responses (common for multimodal providers like Bedrock Titan,
        Nova, Vertex AI), returns the full prompt_tokens_details.
        For multi-item responses, distributes integer fields evenly across items
        so that summing all per-item details reconstructs the original totals.
        """
        if result.usage is None or result.usage.prompt_tokens_details is None:
            return None

        details: Final = result.usage.prompt_tokens_details

View on GitHub (pinned to 6c2dcb801b)

Solutions

1. Check the interpolated key name in the message to see which field is missing from the embedding response data.
2. Upgrade LiteLLM — provider transformation/cache extraction mismatches are usually fixed in patch releases.
3. As a workaround, disable caching for that embedding model (or route around the cache) until compatible.
4. File an upstream issue with the provider name and response sample.

Example fix

# before
litellm.cache = litellm.Cache(type="redis")
litellm.embedding(model="provider/embed-model", input=["hi"])  # -> ValueError: Missing expected key

# after (temporarily bypass cache for that model)
litellm.embedding(model="provider/embed-model", input=["hi"], cache={"no-cache": True})

def embedding_response_is_well_formed(resp) -> bool:
    try:
        return all(d.get("embedding") is not None and d.get("index") is not None for d in resp.data)
    except (AttributeError, KeyError):
        return False

def is_cacheable_embedding_response(resp) -> bool:
    return hasattr(resp, "data") and all(isinstance(d, dict) or hasattr(d, "embedding") for d in getattr(resp, "data", []))

try:
    resp = litellm.embedding(model=embed_model, input=inp)  # cache enabled
except ValueError as e:
    if "Missing expected key in embedding response" in str(e):
        with cache_disabled():
            resp = litellm.embedding(model=embed_model, input=inp)
    else:
        raise

When it happens

Trigger: Using response caching with embeddings where the provider's EmbeddingResponse data items lack standard keys (e.g. no 'embedding' or 'index'), so the dict/vars-based extraction in the caching helper raises KeyError, which is caught and re-raised as this ValueError.

Common situations: A new or non-standard embedding provider whose transformation returns incomplete data objects; a LiteLLM version where a provider transformation was updated but the caching extraction was not; multimodal providers with atypical per-item structure.

Related errors

• Setting user/encoding format is not supported by {custom_llm
• input must be a string or a list
• LLM client cache lazy import: unknown attribute {name!r}
• WXO: No run_id in response: {run_data}
• cache key is None

AI-assisted analysis of BerriAI/litellm@6c2dcb801b (2026-08-15). Data as JSON: /api/errors/1eb5153f7e148967. Report an issue: GitHub.