Allow to override `retrieval_chain` flag (for custom chains with multi-key output)

[nb-programmer]

Title says all. Current implementation assumes len(chain.output_keys) > 1 is a retrieval chain, but that is not always the case, since it can be any other type of chain (like a custom one which won't have the source key).

Or alternatively, the detection may be made more strict by checking the actual output keys

[msoedov]

Hi @nb-programmer, do you have a code example that reproduces this issue?

[nb-programmer]

Sure:

demo_bug.py

from langchain.chains.base import Chain
from langchain.callbacks.manager import CallbackManagerForChainRun

from typing import Dict, Optional, Any, List


class CustomChain(Chain):
    input_key: str = "input"
    output_key: str = "output"

    def _call(
        self,
        inputs: Dict[str, Any],
        run_manager: Optional[CallbackManagerForChainRun] = None,
    ) -> Dict[str, str]:
        print("input:", inputs)
        return {self.output_key: "Hello", "other": "test"}

    @property
    def input_keys(self) -> List[str]:
        return [self.input_key]

    @property
    def output_keys(self) -> List[str]:
        """:meta private:"""
        return [self.output_key, "other"]


chain = CustomChain()

$ langcorn server demo_bug:chain

Then call the endpoint with any input, eg:

{
  "input": "test"
}

500 Error: Internal Server Error

File "venv\lib\site-packages\langcorn\server\api.py", line 118, in handler
    source_documents=[str(t) for t in output.get("source_documents")],
TypeError: 'NoneType' object is not iterable

Expected: It should return all keys, or just the output key without exception

[msoedov]

Thx for the example @nb-programmer . Fixed the bug in e803e5d . Going to release a new version today