Reason across everything your enterprise knows with proof.
cognexas 2.0 combines a domain-tuned retrieval model, a fine-tuned cross-encoder and an interactive entity graph into one platform. Every answer ships with its sources, its confidence, and the six-step trace that produced it.
- 1Intent parse
- 2Retrieve & score
- 3Rank & filter
- 4Cross-reference
- 5Synthesise
- 6Confidence
Engineering in Colombo · Enterprise support in San Francisco · Deployed across 12+ countries
The engine, running in production.
Across every deployed workspace averages, not benchmarks. Numbers refresh every 20 seconds.
Your enterprise already knows the answer. It just can't find it.
Knowledge is fragmented across wikis, drives, PDFs and inboxes and the tools that promise to search it can't explain themselves.
Siloed knowledge
The policy that governs a vendor lives in Legal. The runbook that onboards them lives in IT. Nobody connects the two.
Untraceable answers
Black-box assistants produce confident answers with no source, no audit trail and no way to defend the decision.
Slow decision cycles
Analysts spend hours hunting for precedent instead of acting on it. Every repeat question costs the same again.
Four pillars, one reasoning surface
From ingestion to explanation cognexas covers the whole knowledge lifecycle.
Knowledge Ingestion
Policies, contracts, runbooks, tickets and research notes normalised into one indexed corpus.
Reasoning Engine
Weighted retrieval, entity cross-referencing and confidence scoring on every query.
Explainability Trace
Every answer ships with its numbered reasoning steps and the exact sources that produced it.
Enterprise Governance
Role controls, immutable audit trail, domain scoping and on-prem or air-gapped deployment.
Ingest. Model. Reason. Act.
A deterministic pipeline you can inspect at every stage no opaque model weights.
Ingest
Documents are embedded by the custom 340M encoder, tagged with type, domain and entities, then added to the HNSW index.
Model
Entities and their relationships are woven into a typed graph with 9 node classes and weighted edges.
Reason
Six ordered steps: intent parse → retrieve → rank & filter → cross-reference → synthesise → confidence.
Act
The chain, sources and confidence land in the console. Every step is replayable, every source citable.
Not an update. A rebuild.
Every subsystem was rewritten retrieval, ranking, synthesis, graph traversal and the audit layer. The measured results below come from the same 12-document corpus running on identical hardware before and after the rewrite.
10× faster retrieval
Retrieval was reimplemented around a custom embedding index with an HNSW graph, replacing the linear keyword scan of v1.x. Scoring is now constant-time per document.
Recall@10 · 68% → 94%
A domain-tuned embedding model, trained on internal legal, risk and IT corpora, replaces the bag-of-words scorer. Reranker is a fine-tuned cross-encoder.
Pluggable synthesis
Step 5 of the reasoning chain is now an interface. Template synthesis ships by default; an LLM backend can be attached without touching the trace, sources or confidence.
Per-step confidence
Each of the six reasoning steps reports its own contribution to the final score, calibrated against a held-out audit set. Brier score improved from 0.21 to 0.08.
Interactive entity explorer
Force-directed graph with zoom, pan, typed clustering and shortest-path search. Entity types expanded from 4 to 9.
SSO · RBAC · audit v2
SAML 2.0 and OIDC with SCIM provisioning. Role-based access scoped per domain, workspace and document type. Immutable audit log.
cognexas 1.x
- Keyword-weighted linear scan
- 68% recall@10
- 1.9 s p50 retrieval
- 3-step reasoning chain
- Single-source answers
- 4 entity types · text list
- Stub authentication
- 30-day audit retention
- Single-region deploy
cognexas 2.0
- Custom embedding index + HNSW
- 94% recall@10
- 180 ms p50 retrieval
- 6-step chain · per-step confidence
- Multi-source synthesis (3–5 sources)
- 9 entity types · interactive graph
- SAML 2.0 · OIDC · SCIM · RBAC
- 1-year audit retention (default)
- On-prem · air-gapped · multi-region
Trained on NVIDIA. Served on your hardware.
cognexas 2.0 ships with custom-trained models rather than calling out to a third-party API. Predictable latency, no per-query fees and full data residency.
Custom embedding model
A 340M-parameter encoder fine-tuned on legal, risk and IT corpora. Trained on NVIDIA A100 80GB GPUs over three weeks on ~2.1M documents.
Fine-tuned cross-encoder reranker
Top-K results are re-scored by a cross-encoder trained on human-annotated relevance judgements. Lifts recall@10 from 68% to 94%.
TensorRT-LLM inference
Inference is served through NVIDIA TensorRT-LLM on L40S GPUs, with an ONNX Runtime fallback for CPU-only air-gapped deployments.
Vector index · HNSW
Embeddings live in a hierarchical navigable small-world index, rebuilt incrementally on every document change. 12M vectors per node at <2 GB resident memory.
Confidence calibration
Per-step and final confidence scores are calibrated on a held-out audit set of 4,800 query-answer-source triples. Brier 0.21 → 0.08.
Pluggable LLM synthesis
Step 5 accepts any OpenAI-compatible endpoint public API, self-hosted vLLM, or your own fine-tuned model. Trace, sources and confidence unchanged.
Deployment stack
NVIDIA A100 80GB × 8 · PyTorch · DeepSpeed · custom data pipeline
NVIDIA TensorRT-LLM · Triton Inference Server · ONNX Runtime (CPU fallback)
PostgreSQL + pgvector · custom HNSW index · encrypted at rest (AES-256)
Kubernetes · Helm charts · air-gapped registry mirror · ArgoCD
Everything you need to trust an answer
Entity Graph
Nine node types, weighted edges, path finding and typed clustering all in the browser.
Source Citations
Every claim links to the document and relevance score that produced it.
Confidence Scoring
Per-step and final confidence, calibrated to a Brier score of 0.08 on a held-out audit set.
Audit Log
Every create, update, query and deletion recorded with actor, timestamp and SHA-256 payload digest.
Domain Filters
Scope a query to risk, legal, IT or R&D to cut noise at retrieval time.
Role Controls
SAML 2.0, OIDC and SCIM. Roles scoped per domain, workspace and document type.
On-Prem Deployment
Helm charts, air-gapped registry mirror, ONNX Runtime CPU fallback for offline clusters.
REST API
Every console action is a documented REST endpoint you can wire into your own systems.
Replayable History
Re-run any past query and get the identical trace same sources, same confidence, same answer.
Everything you need, on one screen.
This is the real console. Click around the sidebar. Ask a question. Watch the six-step chain run.
-
reason.queryvendor onboarding · 82% confidence2m
-
knowledge.createadded "API Gateway Architecture Standard"18m
-
auth.logindemo@cognexas.net1h
| Document | Type | Domain |
|---|
One endpoint. Every action.
The same reasoning engine that powers the console is exposed as a documented REST API. Call it from your own systems, run it in CI, or wire it into Slack.
curl -X POST https://api.cognexas.net/v2/reason \
-H "Authorization: Bearer cx_live_…" \
-H "Content-Type: application/json" \
-d '{
"query": "What must a Tier 1 vendor provide?",
"domain": "risk",
"top_k": 5
}'import { cognexas } from "@cognexas/sdk";
const cx = cognexas(process.env.CX_KEY);
const r = await cx.reason({
query: "What must a Tier 1 vendor provide?",
domain: "risk",
topK: 5
});
console.log(r.answer);
console.log(r.sources);from cognexas import Client
cx = Client(api_key=os.environ["CX_KEY"])
r = cx.reason(
query="What must a Tier 1 vendor provide?",
domain="risk",
top_k=5,
)
print(r.answer)
for s in r.sources:
print(s.title, s.score){
"ok": true,
"answer": "Tier 1 vendors require a full security…",
"confidence": 0.82,
"latency_ms": 174,
"steps": [
{ "n": 1, "label": "Intent parse" },
{ "n": 2, "label": "Retrieve & score" },
{ "n": 3, "label": "Rank & filter" },
{ "n": 4, "label": "Cross-reference" },
{ "n": 5, "label": "Synthesise" },
{ "n": 6, "label": "Confidence" }
],
"sources": [
{ "id": "kb_001", "score": 0.87 },
{ "id": "kb_012", "score": 0.61 }
]
}Deploy it where your data already lives
cognexas runs inside your own perimeter. Nothing leaves the network unless you configure an egress. Air-gapped deployments run with no outbound traffic at all.
- SAML 2.0 · OIDC · SCIM 2.0 · RBAC scoped per domain
- Immutable audit trail · SHA-256 payload digest · 1-year retention default
- AES-256 at rest · TLS 1.3 in transit · mTLS between services
- SOC 2 Type II audit in progress · ISO 27001 controls mapped
- Air-gapped edition · signed offline model bundles · no egress
Runs where your data already lives.
The same engine, three deployment shapes. Pick the one that matches your compliance envelope you can change it any time.
Cloud
Fully managed. Multi-region. Zero-ops.
- AWS · GCP · Azure marketplace
- 4 global regions
- Deploy in under 5 minutes
- Managed backups & updates
On-Prem
Your VPC. Your hardware. Your rules.
- Kubernetes · Helm · Terraform
- Runs on L40S · A100 · A10
- mTLS between every service
- 1-year audit retention default
Air-Gapped
Zero network access. Signed bundles only.
- No outbound egress, ever
- Signed offline bundles
- ONNX CPU fallback available
- Dedicated solution engineer
From a search tool to a reasoning platform.
Three years, three product generations and one conviction: enterprises don't need another search box they need a system that reasons and can prove it.
Founded in Colombo
Arjun Fernando and a team of three start cognexas after two years of internal research at a large Sri Lankan bank.
Companycognexas 1.0 ships to first customer
Deterministic keyword-weighted retrieval. Reasoning chain of three steps. First deployment at Northwind's risk team, on-prem.
ProductSeed round · team of eight
Entity graph v1 (four types, ~40 relations). SSO and RBAC land. First air-gapped deployment at a government-adjacent research lab.
Company · ProductModel training begins on NVIDIA infrastructure
Custom embedding and reranker models trained on a cluster of NVIDIA A100 80GB GPUs. First production inference served via TensorRT-LLM on L40S GPUs.
AI / MLAir-gapped edition · SOC 2 Type II in progress
Fully offline deployment validated at two regulated customers. Independent audit engagement begins. Audit trail extended to 12 months by default.
Compliancecognexas 2.0 GA
Rewritten reasoning engine. Retrieval latency drops from 1.9 s to 180 ms. Six-step trace with per-step confidence. Interactive entity graph with path finding. Pluggable synthesis for LLM integration.
LiveReasoning graphs v2
Multi-hop reasoning across documents with evidence chains longer than a single retrieval pass. Formal verification of chain integrity for audit use cases.
RoadmapCustom model marketplace
Customers train domain-specific retrieval models on their own corpus, using the same NVIDIA infrastructure pipeline that ships with the platform.
RoadmapWe built cognexas because the enterprises we worked with didn't need another search box. They needed a system that could look at a policy, a contract, a runbook and a research note and tell them, with citations, what to do next. Version 2.0 is the platform we wished existed when we started: custom models, measurable retrieval quality, and an audit trail that survives a regulator's questions.
What changes when answers explain themselves
"The reasoning chain ended the debate. We can show an auditor exactly which clause produced the answer and how we got there with the same confidence score we saw at query time."
"2.0 cut our query time from two seconds to under a quarter. Our analysts stopped hunting through folders and started trusting the system the numbers are on the dashboard."
"Running it on-prem with no outbound egress was non-negotiable. cognexas shipped the air-gapped edition in our VPC in under a week signed bundles, no cloud calls, ever."
Start small, scale to air-gapped
Every plan includes the full reasoning engine, custom models and audit trail. No feature gating on explainability.
Starter
For a single team proving the value.
- Up to 500 documents
- Full 6-step reasoning chain
- 3 domains · 3 entity types
- 30-day audit retention
Enterprise
For cross-functional knowledge programmes.
- Unlimited documents
- Custom embedding + reranker models
- Interactive entity graph · 9 types
- SAML · OIDC · SCIM · RBAC
- Pluggable LLM synthesis
- 1-year audit retention
Air-Gapped
For regulated and classified environments.
- Fully offline · no egress
- Bring-your-own NVIDIA hardware
- Signed offline model bundles
- Dedicated solution engineer
- Custom retention policy
Questions, answered
Five measurable things. Retrieval latency dropped from 1.9 s to 180 ms p50. Recall@10 improved from 68% to 94%. The reasoning chain grew from 3 to 6 steps, each reporting its own confidence. The entity graph gained 5 new node types and became interactive. And synthesis became pluggable you can attach any OpenAI-compatible LLM without losing the trace or the citations.
Yes. The embedding model (340M params, 1024-dim) and the cross-encoder reranker were both trained on NVIDIA A100 80GB GPUs against ~2.1M documents and 180k annotated relevance pairs. They ship with the platform as signed artifacts and run locally on your hardware no per-query API calls, no data leaving your perimeter.
Only if you want it to. By default, step 5 (synthesis) uses template-based generation, which means answers are fully reproducible and deterministic. In 2.0 you can attach an LLM backend OpenAI, Anthropic, or a self-hosted vLLM instance for more natural prose, while the trace, sources and confidence scoring stay identical.
The cluster has no outbound network access at all. Model updates, index rebuilds and even license validation are handled through signed offline bundles that you install manually. This is the same deployment pattern used by defense and regulated-finance customers.
Sri Lankan customers contract with Cognexas Innovations (Pvt) Ltd (18 Flower Road, Colombo 07). Customers in the United States and the rest of the world contract with Cognexas Innovations LLC (600 Montgomery Street, Suite 1800, San Francisco, CA 94111).
Two announced milestones. Q3 2026: reasoning graphs v2 multi-hop evidence chains spanning multiple documents, with formal verification of chain integrity. Q1 2027: a custom model marketplace you train domain-specific retrieval models on your own corpus using the same NVIDIA pipeline that ships with the platform.
Most teams have their first domain indexed and answering questions within a week. Ingestion connectors for wikis, drives and ticketing systems ship out of the box. Air-gapped deployments typically take two to three weeks because of the offline bundle workflow.
The console surfaces it. Below your configured threshold, cognexas tells you it could not find sufficient evidence rather than inventing an answer and shows you what it did find. Per-step confidence lets you see which step of the chain weakened the result.
Talk to the team behind the platform.
Tell us what you need to reason over a policy set, a contract archive, a runbook library and we'll show you exactly how cognexas handles it.