LlamaIndex

LlamaIndex can route generation and embeddings through separate compatible adapters. Configure both; changing only the language model can leave embeddings pointing at a default provider.

1. Install adapters

python -m pip install -U llama-index-core llama-index-llms-openai-like llama-index-embeddings-openai-like

Set FENNO_API_KEY, FENNO_MODEL, and FENNO_CONTEXT_WINDOW to the chosen chat model’s verified context size as an integer. Set FENNO_EMBEDDING_MODEL separately if using retrieval. Pin validated package versions for deployment.

2. Verify the language model

import os
from llama_index.llms.openai_like import OpenAILike
 
llm = OpenAILike(
    model=os.environ["FENNO_MODEL"],
    api_base="https://api.fenno.ai/v1",
    api_key=os.environ["FENNO_API_KEY"],
    context_window=int(os.environ["FENNO_CONTEXT_WINDOW"]),
    is_chat_model=True,
    is_function_calling_model=False,
    max_retries=0,
    timeout=60,
)
print(llm.complete("Reply with OK only."))

api_base is LlamaIndex’s option name, unlike base_url in the OpenAI Python client. is_chat_model=True selects chat behavior. The example disables function calling because a simple text request does not require it; only enable it after verifying model support.

3. Configure embeddings before indexing

Continue in the same Python process with an available embedding model:

from llama_index.core import Document, Settings, VectorStoreIndex
from llama_index.embeddings.openai_like import OpenAILikeEmbedding
 
Settings.llm = llm
Settings.embed_model = OpenAILikeEmbedding(
    model_name=os.environ["FENNO_EMBEDDING_MODEL"],
    api_base="https://api.fenno.ai/v1",
    api_key=os.environ["FENNO_API_KEY"],
    max_retries=0,
)
documents = [Document(text="Our support team works Monday to Friday, 09:00 to 18:00.")]
index = VectorStoreIndex.from_documents(documents)
answer = index.as_query_engine().query("When is the support team available?")
print(answer)

This synthetic document is an example, not Fenno’s support schedule. Index construction sends document text to the embedding endpoint; querying sends query text for embedding and retrieved context to the chat endpoint.

Production boundaries

• Never mix vectors from different models or dimensions in the same index.
• Scope retrieval to the current user’s document permissions before building the model context.
• Settings is process-wide configuration. For multi-tenant or concurrent applications with different providers, prefer explicit per-component clients instead of mutating global settings per request.
• The framework does not provide a hosted Fenno knowledge base, vector database or guaranteed factual answers.
• A custom model ID can require explicit metadata; do not pass it to a provider-specific adapter that assumes a fixed official model catalogue.
• Keep tool permissions, callback/tracing data handling and retries under application control.

See the official OpenAILike reference , OpenAILikeEmbedding reference , and retrieval guide.