Model outputs gibberish using Local and Global search.

[Tejaswgupta]

I got correct result in the first iteration of playing around with the dickens book text. However when i tried with some other long form texts , the naive type query worked fine while other took too long and also output gibberish.

I'm using Qwen 2.5 7B/14B INT4 models with 32k context length which are hosted with vllm on a VM.

Logs: https://pastebin.com/fvy3DsSH

Code snippet:

WORKING_DIR = "./dickens"

if not os.path.exists(WORKING_DIR):
    os.mkdir(WORKING_DIR)

async def llm_model_func(
    prompt, system_prompt=None, history_messages=[], **kwargs
) -> str:
    return await openai_complete_if_cache(
        # "gemini-flash",
        "self-hosted-small", #this is the qwen-2.5-7b-gptq-int4 model
        prompt,
        system_prompt=system_prompt,
        history_messages=history_messages,
        api_key="emtp",
        base_url="https://xx.xx.com",
        **kwargs,
    )

async def embedding_func(texts: list[str]) -> np.ndarray:
    return await openai_embedding(
        texts,
        model="embeddings-finetune/gte-base-case-law-v2",
        api_key="emtp",
        base_url="https://xx.xx.com",
    )

async def get_embedding_dim():
    test_text = ["This is a test sentence."]
    embedding = await embedding_func(test_text)
    embedding_dim = embedding.shape[1]
    return embedding_dim

# function test
async def test_funcs():
    result = await llm_model_func("How are you?")
    print("llm_model_func: ", result)

    result = await embedding_func(["How are you?"])
    print("embedding_func: ", result)

# asyncio.run(test_funcs())

async def main():
    try:
        embedding_dimension = await get_embedding_dim()
        print(f"Detected embedding dimension: {embedding_dimension}")

        rag = LightRAG(
            working_dir=WORKING_DIR,
            llm_model_func=llm_model_func,
            embedding_func=EmbeddingFunc(
                embedding_dim=embedding_dimension,
                max_token_size=8192,
                func=embedding_func,
            ),
        )

        with open("./book.txt", "r", encoding="utf-8") as f:
            await rag.ainsert(f.read())

        # Perform naive search
        print(
            await rag.aquery(
                "What are the top themes in this story?", param=QueryParam(mode="naive")
            )
        )
        print('---'*50)

        # Perform local search
        print(
            await rag.aquery(
                "What are the top themes in this story?", param=QueryParam(mode="local")
            )
        )
        print('---'*50)

        # Perform global search
        print(
            await rag.aquery(
                "What are the top themes in this story?",
                param=QueryParam(mode="global"),
            )
        )
        print('---'*50)

        # Perform hybrid search
        print(
            await rag.aquery(
                "What are the top themes in this story?",
                param=QueryParam(mode="hybrid"),
            )
        )
    except Exception as e:
        print(f"An error occurred: {e}")

[nihirv]

Can you give an example of what gibberish means? Literal gibberish (i.e. strings of tokens that make no sense), or coherent text but unrelated to the source documents?

[Tejaswgupta]

@nihirv you can check the logs i've shared. It's literal gibberish.