LangExtract failing with ResolverParsingError: Content must contain an 'extractions' key (worked before) Body:

[kmeersmans]

Hello!

I’m running into a persistent issue with Google’s langextract that started recently.
The same code and prompts worked before, but now consistently fail.

ResolverParsingError: Content must contain an 'extractions' key.

This indicates that the model output does not match the schema expected by LangExtract (i.e. the response JSON does not contain an extractions key).

What I observe

LangExtract runs, the LLM responds, but

• the output is valid text / JSON-like
• However, it does not include an extractions field
• The resolver then crashes when parsing the response

Models tested

• gemma
• llama

Same behavior on all models.

Hass anyone run into the same issues, and how can it be solved?

Cheers,
Karen

[kastnerp]

Looks like setting max_char_buffer=500 (or something smaller than default) is a workaround on v1.1.1

[kmeersmans]

Sadly, this doesn't work for me :( Would you mind sharing the other parameters so I can compare all of them?

[kastnerp]

Actually, this is not the solution. Could it be that your prompt has some instructions that urges the model to overwrite the dictionary output? For me this was the case and fixed the issue.

[kastnerp]

LangExtract: Processing, current=83,329 chars, processed=0 chars:  [00:44]
⚠️ Attempt 1/2 failed for 66DP2YUU_-_Garc_a-Mart_n_-_2025_-_Perceived_constraints_for_psyc.json: Content must contain an 'extractions' key.
   ↪ Retrying with use_schema_constraints=False, fence_output=False...
   ⏳ Extracting with LLM (Attempt 2)...

So for me, this happens when I try with models that are not llama or gemma, such as qwen3:14b. Would anyone know if different configurations are needed for other vendor models?

[Yash-xoxo]

This looks like a model instruction-following issue. A few things to check:

1. Model compatibility
Some models struggle with structured output. From your logs, qwen3:14b is failing. Try:

• llama3.3 or llama3.1 (best for JSON)
• gemma2:27b (more reliable than smaller gemmas)
• Increase temperature to 0 for deterministic output

2. Prompt conflict
@kastnerp's point is key - check if your prompt has instructions like:

"Return only the answer"
"Format as markdown"
"Respond concisely"

These can override the schema. Remove any output format instructions from your prompt.

3. Schema enforcement
The retry with use_schema_constraints=False suggests the model can't follow the schema. Try forcing it:

extractor = LangExtract(
    model="llama3.3",
    use_schema_constraints=True,
    fence_output=True,  # Wrap in ```json
    max_char_buffer=1000  # Smaller chunks
)

4. Debug the raw output
Add logging to see what the model actually returns:

# Patch temporarily to inspect
import langextract
original_parse = langextract.resolver._parse_response

def debug_parse(response):
    print(f"Raw response: {response[:500]}")
    return original_parse(response)
    
langextract.resolver._parse_response = debug_parse

For qwen specifically, you might need to adjust the system prompt template - it uses different instruction formats than Llama.