Knowledge Graph AI in India

Enterprises in BFSI, legal, and manufacturing generate large volumes of unstructured Indian business documents — contracts, filings, correspondence — that need to become a structured knowledge graph before they're actually useful for search or analytics. Large language models can do much of the entity and relation extraction that used to require custom NLP pipelines, and unoblox bills that work per million tokens, in rupees, through one endpoint.

Where LLMs help build a knowledge graph

The core of any knowledge graph is entities (people, companies, contracts, assets) and the relations between them (owns, is-a-party-to, is-subsidiary-of). LLMs are useful for pulling both out of unstructured text in a consistent, structured format — typically JSON — that a graph database can ingest directly. They're also useful for a lighter task: suggesting when two extracted entities likely refer to the same real-world thing, ahead of a proper entity-resolution step.

Model choices by extraction volume

Workload	Suggested model	₹ input / ₹ output (per 1M tokens)
High-volume bulk extraction	deepseek-ai/deepseek-v4-flash	₹9.07 / ₹18.14
Structured JSON extraction at moderate cost	qwen/qwen3-235b-a22b-instruct-2507	₹9.07 / ₹55.44
Complex multi-hop relation inference	openai/gpt-5	₹126 / ₹1008
Prototyping your extraction schema	qwen/qwen3-1.7b	FREE (₹0)

A sample extraction call

curl https://api.unoblox.ai/v1/chat/completions \
  -H "Authorization: Bearer ub-gw-xxxxxxxxxxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3-235b-a22b-instruct-2507",
    "messages": [{"role": "user", "content": "Extract all companies and their relationships from this filing as a JSON list of source, relation, target."}]
  }'

Embeddings for entity linking

Once you've extracted candidate entities, unoblox's native /v1/embeddings endpoint can help flag likely duplicates by similarity before a human or a rules-based step confirms the match — useful when the same company appears under slightly different names across documents. The specific embedding model and its rate can change as the catalog grows, so check /models for the current one rather than assuming a fixed price.

Getting extraction accuracy right

No extraction pipeline is perfect on the first pass, and no model here should be trusted to build a production knowledge graph unsupervised. Start with the cheap, fast model on a sample set, have a human review a portion of the output, and only move to a stronger (and pricier) model like openai/gpt-5 for the relations that the cheap pass gets wrong or flags as low-confidence.

Frequently asked questions

Can an LLM build my entire knowledge graph unsupervised? Not reliably — treat it as an extraction assistant that needs a validation step, not a fully automated pipeline on day one.

Which model should handle the first extraction pass? A cheap, fast model like deepseek-ai/deepseek-v4-flash on a representative sample, so you can measure accuracy before committing to volume.

Does unoblox include a graph database? No — unoblox provides the language models and embeddings for extraction and linking; you still need your own graph store to hold the result.

Is there a free way to test extraction prompts? Yes, qwen/qwen3-1.7b is free (₹0).

Is pricing in rupees for embeddings too? Yes — every endpoint on unoblox, including embeddings, is priced per million tokens in rupees.

Do I need separate contracts for the chat and embedding models? No — one ub-gw-... key and one monthly GST invoice cover every endpoint.

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