LangChain with an India AI API (₹ billing)

LangChain is the dominant agent and RAG framework in Python and JavaScript. unoblox makes it ₹-native: point LangChain at unoblox instead of OpenAI, choose any model — Qwen for cost, Claude for nuance, DeepSeek for reasoning, GPT-4o for general use — and get one GST invoice in rupees.

Why unoblox + LangChain

• One-line setup: swap base_url and api_key, keep every chain and prompt template as-is
• 40+ models: Qwen3.8-27B for everyday RAG (₹16.32/1M input), Qwen3 1.7B free for prototyping, Claude Opus for hard reasoning (₹504/1M input)
• No rewrite: LangChain's interfaces are unchanged; only configuration differs
• ₹ billing: one monthly GST invoice, input-tax-credit eligible, no currency exposure
• Both languages: works with LangChain Python and LangChain.js identically

Two-line integration

# Modern LangChain (langchain_openai package)
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
  api_key="ub-gw-YOUR_API_KEY",
  base_url="https://api.unoblox.ai/v1",
  model="openai/gpt-4o"
)

# Use LangChain exactly as documented
chain = prompt | llm | output_parser
result = chain.invoke({"input": "What is the capital of France?"})

All LangChain features — memory, tools, chains, retrieval — work through the unoblox endpoint unchanged.

Model selection for common tasks

LangChain task	Recommended model	₹ per 1M input
Semantic search + chunking	Qwen3.8-27B	₹16.32
RAG retrieval, low cost	Qwen3 1.7B	FREE
Agentic reasoning	Claude Opus	₹504
Fast summarization	DeepSeek V4 Flash	₹9.07
Multi-turn chat	GPT-4o	₹252

Start on Qwen3 1.7B while you build the chain; move a step up only where quality actually stalls.

Full example: RAG chain

from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_community.vectorstores import FAISS
from langchain.chains import RetrievalQA

embeddings = OpenAIEmbeddings(
  api_key="ub-gw-...",
  base_url="https://api.unoblox.ai/v1",
  model="Qwen/Qwen3-Embedding-0.6B"
)

vectorstore = FAISS.from_texts(docs, embeddings)

llm = ChatOpenAI(
  api_key="ub-gw-...",
  base_url="https://api.unoblox.ai/v1",
  model="anthropic/claude-opus-5"
)

chain = RetrievalQA.from_chain_type(llm=llm, retriever=vectorstore.as_retriever())
chain.run("Summarize the documents")

One endpoint, one key, two unoblox models — embeddings free, the answering LLM at ₹504 per 1M input tokens.

Troubleshooting

401 Unauthorized when calling unoblox? Check that api_key matches your dashboard value. unoblox keys are prefixed ub-gw-; OpenAI keys start with sk- — a leftover OpenAI key is the usual cause.

"Model not found" error? Query https://api.unoblox.ai/v1/models to see the live catalog. Model ids are case-sensitive — most chat models are lowercase, but the embedding model is Qwen/Qwen3-Embedding-0.6B with a capital Q.

Slow embeddings? Qwen3-Embedding-0.6B is a small model; expect sub-second inference per chunk. Embed once and cache the vectors instead of re-embedding on every query.

How do I monitor costs? Export usage logs from your unoblox dashboard (90–180 days retained). Every call logs model name, tokens, and ₹ cost.

Can I use LangChain.js (Node.js)? Yes. npm install @langchain/openai. Configuration is identical — one base_url, one key, the same model ids.

Do tool-use chains work? Yes. LangChain's tool-calling agents work with Claude, GPT-4o, and Qwen — check the model card on /models if a specific model doesn't support tools.

Get started in rupees → https://unoblox.ai/sign-in