vrin 1.3.2

Knowledge reasoning infrastructure for AI — structure documents into a temporal knowledge graph, reason across them, and trace every answer to specific facts

VRIN — Knowledge Reasoning Infrastructure for AI

The reasoning engine for AI that needs to be right.

VRIN structures your documents into a temporal knowledge graph and reasons across them — giving AI answers you can trace to specific facts.

Benchmark	VRIN	Best Baseline	Gap
MultiHop-RAG	95.1%	78.9% (GPT 5.2 w/ same docs)	+16.2pp
MuSiQue	28% more accurate	HippoRAG 2 (academic SOTA)	+10.6 EM
FinQA	71.4%	11.1% (vector-only retrieval)	+640%

Installation

pip install vrin

Quick Start

from vrin import VRINClient

client = VRINClient(api_key="vrin_your_api_key")

# Insert knowledge
client.insert("ACME Corp reported $50M revenue in Q4 2025.", title="ACME Financials")

# Query — answers traced to specific facts
result = client.query("What is ACME's revenue?")
print(result["summary"])

Bulk Ingestion with Adaptive Concurrency

Ingest hundreds of documents with automatic concurrency control. The SDK uses Netflix's Gradient2 algorithm to monitor backend latency and adjust parallelism in real time — ramping up when healthy, backing off when congested, retrying failures with exponential backoff, and running a sequential recovery pass for any remaining items. Zero information loss guaranteed.

items = [
    {"content": "Apple reported $416B revenue in FY2025.", "title": "AAPL 10-K"},
    {"content": "Microsoft Cloud surpassed $50B quarterly.", "title": "MSFT Earnings"},
    # ... hundreds more
]

result = client.bulk_insert(items)
print(f"Ingested {result['completed']}/{result['total']} — "
      f"{result['facts_stored']} facts stored")

Streaming

for token in client.query("Summarize Q4 results", stream=True):
    print(token, end="", flush=True)

Query Modes

result = client.query(
    "Compare ACME and Widget Corp revenues",
    response_mode="research",     # "chat" | "thinking" | "research"
    query_depth="research",       # "basic" | "thinking" | "research"
)

File Upload

client.upload_file("report.pdf", save_to_memory=True)

Conversations

client.start_conversation()
r1 = client.continue_conversation("What was ACME's Q4 revenue?")
r2 = client.continue_conversation("How does that compare to Q3?")  # has context
client.end_conversation()

MCP Integration

VRIN exposes an MCP (Model Context Protocol) server so any compatible AI assistant can query your knowledge base — Claude Code, Claude Desktop, Cursor, Windsurf, or custom agents.

Tools exposed: vrin_query_async, vrin_check_job, vrin_search_entities, vrin_get_facts

Enterprise — Your Data Stays in Your Cloud

Enterprise API keys (vrin_ent_*) route all data through your own AWS/Azure account. Your knowledge graph, your vector store, your encryption keys. Data never touches our infrastructure.

from vrin import VRINEnterpriseClient

client = VRINEnterpriseClient(api_key="vrin_ent_your_key")
result = client.query("What is our Q4 revenue?")

Three deployment modes: VRIN Cloud, Hybrid Cloud (your data, our compute), Private VPC (everything in your account).

What Makes VRIN Different

	Vector-Only Retrieval	VRIN
Answers	Similar-looking text chunks	Specific facts, traced to sources
Temporal	None — returns latest by similarity	Bi-temporal versioning: "What was true in Q3?"
Cross-document	Concatenates chunks	Traverses entity relationships across documents
Numbers	LLM interprets from raw text	Constraint extraction + validated numerical fields
Aggregation	LLM must infer	Explicit sum/count with calculation steps
Audit trail	Chunk-level at best	Fact-level with document, confidence, and provenance

License

MIT License — see LICENSE file for details.

---

Built by the VRIN Team | vrin.cloud | [e-mail]