ERROR:nemoguardrails.server.api:Access to the specified path is not allowed.

[NirnayK]

I am trying to use custom_embedding_search_provider but for some reason it keeps giving me error :

ERROR:nemoguardrails.server.api:Access to the specified path is not allowed.
Traceback (most recent call last):
File "/nemoguardrails/nemoguardrails/server/api.py", line 273, in chat_completion
llm_rails = _get_rails(config_ids)
File "/nemoguardrails/nemoguardrails/server/api.py", line 230, in _get_rails
raise ValueError("Access to the specified path is not allowed.")
ValueError: Access to the specified path is not allowed.

I tried it with the exact same config as to https://github.com/NVIDIA/NeMo-Guardrails/tree/develop/tests/test_configs/with_custom_embedding_search_provider but it keeps giving me the same issue.

[Pouyanpi]

Hi @NirnayK,

I assume you are using following command:

nemoguardrails server --config=path/to/config

Can you share with me which version of nemoguardrails you are using? I think using the absolute path to the config would resolve this issue.

[NirnayK]

I am using the latest commit of the develop branch. I guess version 9.0.
This works : (older)
full_path = os.path.normpath(os.path.join(app.rails_config_path, config_id))
if not full_path.startswith(app.rails_config_path):
raise Exception("Not allowed.")

This doesnt: (newer/current)
base_path = os.path.abspath(app.rails_config_path)
full_path = os.path.normpath(os.path.join(base_path, config_id))

    if not full_path.startswith(base_path + os.sep):
        raise ValueError("Access to the specified path is not allowed.")

Lets say i give path to config = folder
Now whether i add a / or not the newer code always appends it in the end and it returns false.
Also wouldnt absolute path to config mess with kb ?

[NirnayK]

Also I shifted to 8.x and tried out the config works but it never calls the custom provider. The init function is called but the search function is never called when user sends a query

[NirnayK]

To sum up the ABC bot works, though I am not sure about the function call flow. However Custom Provider does not work for some reason even though I have followed the doc

[NirnayK]

My config.py :
import os
from typing import List
from e2enetworks.cloud import tir
from qdrant_client import QdrantClient
from nemoguardrails import LLMRails
from nemoguardrails.embeddings.index import EmbeddingsIndex, IndexItem

class CustomEmbeddingSearchProvider(EmbeddingsIndex):
def init(self):
self.host = os.getenv("QDRANT_HOST")
self.key = os.getenv("QDRANT_API_KEY")
self.collection = os.getenv("QDRANT_COLLECTION_NAME")

    assert self.host, "QDRANT URL is not set"
    assert self.key, "QDRANT API Key is not set"
    assert self.collection, "QDRANT Collection Name is not set"

    self.qdrant_client = self.get_qdrant_client()


    tir.init()
    self.tir_client = tir.ModelAPIClient()

@property
def embedding_size(self):
    self.embedding_size = 4096

@property
def cache_config(self):
    pass

async def _get_embeddings(self, texts: List[str]):
    pass

async def add_item(self, item: IndexItem):
    pass

async def add_items(self, items: List[IndexItem]):
    pass

async def build(self):
    """Build the index, after the items are added.

    This is optional, might not be needed for all implementations."""
    pass

def get_custom_embedding(self, text: str) -> List[float]:

    data = {"prompt": text}
    try:
        response = self.tir_client.infer(model_name="e5-mistral-7b-instruct", data=data)
        result = response.get("outputs",[]).get("data",[])
        return result
    except Exception as e:
        print(f"Error while getting custom embedding: {e}")
        return []

# def chunk_text(text: str):
#     text_splitter = SentenceTransformersTokenTextSplitter()
#     chunks = text_splitter.split_text(text)
#     return chunks

def get_qdrant_client(self) -> QdrantClient:
    qdrant_client = QdrantClient(host=self.host, port=6333, api_key=self.key)
    collection_exists = qdrant_client.collection_exists(self.collection)
    assert collection_exists, f"Collection {self.collection} does not exist"
    return qdrant_client

async def search(self, text: str, max_results: int) -> List[IndexItem]:
    # chunks = self.chunk_text(text)
    # embeddings = [self.get_custom_embedding(chunk) for chunk in chunks]
    embedding = self.get_custom_embedding(text)
    print("Reached search")
    print("Reached search")
    print("Reached search")
    print("Reached search")
    print("Reached search")
    print("Reached search")
    print("Reached search")

    search_result = self.qdrant_client.search(
        collection_name=self.collection,
        query_vector=embedding,
        limit=max_results,
    )

    items = [hit.payload["data"] for hit in search_result]
    result = [IndexItem(text=text,meta = {"body": item["text"]}) for item in items]

    print("returning stuff")
    print("returning stuff")
    print("returning stuff")
    print("returning stuff")
    print("returning stuff")
    print("returning stuff")
    print("returning stuff")
    print(result)

    return result

def init(app: LLMRails):
print("Yes")
print("Yes")
print("Yes")
app.register_embedding_search_provider("custom", CustomEmbeddingSearchProvider)

[drazvan]

Ok, now I understand. You probably don't have any dialog rails in your config, i.e., no "define user xxx". The ABC bot has the "disallowed.co" file. The prompt that is used is the general one, which doesn't use the relevant_chunks. If you add dialog rails, you should see it working. But this is also a bug. We should enable KB support without dialog rails.

[NirnayK]

so i need to add disallowed.co or will adding the following to config.yml solve it :
rails:
input:
flows:
- self check input

output:
flows:
- self check output

dialog:
single_call:
enabled: False

[drazvan]

Yes. I've recorded the issue so we can fix: #556.

[drazvan]

@NirnayK, can you try the following: make sure the folder to which you're pointing the server has a subfolder for the config

/server-config
  /config-X
    config.yml
    rails.co
    config.py
    ...

Maybe there's a bug with the "singe config" mode of the server.

[NirnayK]

Cool tested it out it works