Is paper-qa able to use opencode's big-pickle model?

[66-firebat]

Hi all, I am using opencodes big-pickle model (free inference), and I have been failing to get paper-qa to work with this backend. Here is my code example (the api-key is fake, just put it in as a placeholder)

from paperqa import Settings, ask
from paperqa.settings import AgentSettings

opencode_config = {
    "model_list": [
        {
            "model_name": "big-pickle",
            "litellm_params": {
                "model": "openai/big-pickle",
                "api_base": "https://opencode.ai/zen/v1",
                "api_key": "sk-FAKEAPIKEY",
                "temperature": 0.0,
            },
        }
    ]
}

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(
        llm="big-pickle",
        llm_config=opencode_config,
        summary_llm="big-pickle",
        summary_llm_config=opencode_config,
        agent=AgentSettings(
            agent_type="fake",  # Use fake agent instead (no tool_choice needed)
            agent_llm="big-pickle",
            agent_llm_config=opencode_config,
        ),
        paper_directory="my_papers",
    ),
)

# Use correct attribute name
print(answer_response.session.answer)

[dosubot]

Hi @66-firebat! I'm Dosu and I’m helping the paper-qa team.

@66-firebat Yes, paper-qa can work with custom OpenAI-compatible backends like OpenCode — the configuration approach you're using is on the right track, but there are a couple of important things to check:

1. Use openai_like/ prefix instead of openai/

A breaking change in recent versions of litellm (1.81.7+) requires using the openai_like/ prefix for OpenAI-compatible endpoints rather than openai/ [1]. Your litellm_params.model should be "openai_like/big-pickle" instead of "openai/big-pickle".

2. You're missing embedding configuration

Paper-qa has five distinct LLM configuration points, and any missing one will silently fall back to OpenAI (which will fail without an OpenAI API key) [2]. You need to also configure:

• embedding + embedding_config — for document embeddings
• parsing settings — to disable multimodal enrichment which defaults to gpt-4o

3. Suggested revised configuration:

from paperqa import Settings, ask
from paperqa.settings import AgentSettings, ParsingSettings

opencode_config = {
    "model_list": [
        {
            "model_name": "big-pickle",
            "litellm_params": {
                "model": "openai_like/big-pickle",  # Use openai_like/ prefix
                "api_base": "https://opencode.ai/zen/v1",
                "api_key": "YOUR_API_KEY",
                "temperature": 0.0,
            },
        }
    ]
}

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(
        llm="big-pickle",
        llm_config=opencode_config,
        summary_llm="big-pickle",
        summary_llm_config=opencode_config,
        agent=AgentSettings(
            agent_type="fake",
            agent_llm="big-pickle",
            agent_llm_config=opencode_config,
        ),
        embedding="openai_like/your-embedding-model",  # You'll need an embedding model too
        embedding_config={...},  # Matching embedding config
        parsing=ParsingSettings(
            use_doc_details=False,
            multimodal=False,  # Prevents fallback to gpt-4o for image processing
        ),
        paper_directory="my_papers",
    ),
)

print(answer_response.session.answer)

The agent_type="fake" and answer_response.session.answer parts of your code are correct [3] [4].

One caveat: if OpenCode's big-pickle is a chat/completion model only (not an embedding model), you'll need a separate embedding provider. Could you share the error message you're seeing? That would help narrow down exactly which configuration point is failing.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.