TrailNanonets platform · Context graph

Turn tribal knowledge into one context graph any agent can follow.

Trail is the Nanonets context graph. It turns your policies, procedures, systems, and institutional knowledge into rules an agent can apply, with every rule traceable to the document it came from.

Context graph5 rules · 6 edges
CREDIT ≤ $50KCONTRACT PRICEVOLUME −10%DELIVERY < 5D?+ RUSH SHIP $2KRE-CHECKED EDGE

The problem

Rules collapse when they depend on each other.

As instructions stack up, frontier models silently drop rules, double-count them, or apply them in the wrong order. Agents break completely when rules conditionally activate or override one another.

Watch the same prompt run two ways. Toggle between a flat context and a context graph.

prompt: confirm this PO, 600 units, contract pricing, deliver in 3 days, credit limit $50k

01Resolve credit rule: order total must stay under $50k.
02Apply contract pricing and 10% volume discount (600 units).
03Delivery in 3 days, so add a rush-ship surcharge (+$2k).
04Final total: $51.2k. Credit limit silently exceeded.
The flat model resolves the credit limit, then quietly blows past it two steps later when the rush-ship surcharge fires.

How it works

Four steps, from raw inputs to a verified result.

01Decompose the rules

Parse your documents, policies, and instructions into explicit rules, conditional branches, and approval gates. Every constraint becomes a node, whether it's a credit limit in order management, a GL rule in AP, or a payer requirement in RCM.

Process: PO to confirmed order
CREDIT ≤ $50KGLOBALCONTRACT PRICEVOLUME −10%DELIVERY < 5 DAYS?+ RUSH SHIP $2KDISCOUNT > 15%?STACKS WITHACTIVATESRE-CHECKS CREDITTRIGGERSCREDIT LIMIT RE-CHECKED

02Map the logic between rules

Draw edges between rules that activate, override, or narrow each other. A surcharge that re-checks a limit, a discount that trips an approval gate, a denial code that changes the next step.

Process: PO to confirmed order
CREDIT ≤ $50KGLOBALCONTRACT PRICEVOLUME −10%DELIVERY < 5 DAYS?+ RUSH SHIP $2KDISCOUNT > 15%?STACKS WITHACTIVATESRE-CHECKS CREDITTRIGGERSCREDIT LIMIT RE-CHECKED

03Build graph-guided outputs

This is where the agent takes action. It builds the result step-by-step with active constraints bound to each part, and global rules stay attached as conditions stack up, so they never silently drop midway.

Process: PO to confirmed order
CREDIT ≤ $50KGLOBALCONTRACT PRICEVOLUME −10%DELIVERY < 5 DAYS?+ RUSH SHIP $2KDISCOUNT > 15%?STACKS WITHACTIVATESRE-CHECKS CREDITTRIGGERSCREDIT LIMIT RE-CHECKED

04Check results against the graph

Re-check the finished output against every node and edge. When a late step violates a constraint, the conflict is caught and routed for review before anything is finalized.

Process: PO to confirmed order
CREDIT ≤ $50KGLOBALCONTRACT PRICEVOLUME −10%DELIVERY < 5 DAYS?+ RUSH SHIP $2KDISCOUNT > 15%?STACKS WITHACTIVATESRE-CHECKS CREDITTRIGGERSCREDIT LIMIT RE-CHECKED

Complex Rules benchmark

Trail is the best model for following complex business rules.

The ComplexConstraints benchmark measures performance with densely interlocking rules, where constraints activate, override, and contradict one another. The Trail context graph engine leads the strongest models.

Read the ComplexConstraints benchmark
Task pass rate · top 3 models
1Trail context graph45.0%
2Gemini 3.1 Pro40.4%
3GPT 5.538.7%

Percentage of prompts fully solved on ComplexConstraints. Higher is better.

90.0%
Rubric pass rate

Individual rules satisfied across 1,559 evaluation items in the benchmark.

How it compares

Trail vs the alternatives

Constraint handling
Trail context graph
Every rule is a node; every dependency is an explicit edge.
In-house build
Hand-written if/else rules that grow brittle as policies stack up.
Generic agent platform
Everything in one flat prompt; constraints get silently dropped.
Conditional & overriding rules
Trail context graph
Activate, override, narrow, and contradict edges modeled directly.
In-house build
Each interaction coded by hand, easy to miss a branch.
Generic agent platform
Model double-counts or mis-applies rules that activate each other.
Verification
Trail context graph
Answer re-checked against the graph; violations repaired before output.
In-house build
Manual QA and spot checks after the fact.
Generic agent platform
Hope the model got it right; no structural check.
Traceability
Trail context graph
Every rule cited to the document it came from, on every decision.
In-house build
The rule lives in code, so the reason lives in a commit message.
Generic agent platform
No record of which instruction drove which decision.
Maintenance
Trail context graph
Edit the graph; dependencies propagate automatically.
In-house build
Rules rot as policy changes; every edit is an eng ticket.
Generic agent platform
Re-prompt and pray; regressions are invisible.
Accuracy on entangled constraints
Trail context graph
45.0% task pass, 90.0% rubric pass on ComplexConstraints.
In-house build
Only as good as the branches you remembered to write.
Generic agent platform
40.4% task pass rate (strongest public model).
Time to production
Trail context graph
Deploy on day one; reliability scales with complexity.
In-house build
Weeks of engineering per ruleset.
Generic agent platform
Fast to start, unreliable once rules interlock.

See it run on your process.

From AI-native hypergrowth companies to Fortune 500 enterprises, Nanonets automates business processes so teams can focus on what matters.