cortIQ

Company knowledge as typed, classified, evidence-backed units.

cortIQ is Pyrana's context engine. It turns documents, connectors, conversations, and system data into Context Units in a knowledge graph: each a claim, its evidence passage with source and location, a knowledge type, a classification, an evidence grade, and a quality score, with an identity that is a content hash of claim and evidence. Agents retrieve them in three modes, and every run accounts for what was retrieved, injected, and cited.

Exec read 3 min Full read 10 min

Context Unit

cxu:1220a7ff…f843

Claim

“Intercompany balances between operating companies settle within 45 days of invoice date from 1 July 2026.”

Context passage

“Effective 1 July 2026, intercompany balances shall be settled within 45 days of invoice date.” · treasury-policy-2026.pdf@v4 · p.12 · §4.2

type: prescribed
class: regulated
evidence grade: A · governing policy document
identity: cxu:1220a7ff…f843 · content hash of claim and evidence
supersedes: cxu:8e02…19d1 · net 60 · prior hash carried
quality: 0.91 · scored from use
links: refines cxu:c07d…9e10 (settlement calendar) · same set: intercompany-rules

NoteIllustrativeretrieved ⊇ injected ⊇ cited

01Why a context engine

A type, a class, a grade, and a hash on every claim.

Retrieval over text chunks returns paragraphs sized for an embedding model. A paragraph holds a dozen claims, one of them stale, and a citation to it proves little. A Context Unit is one claim with its evidence passage, its source and location, and an identity that changes when the claim or the evidence changes. Because citations are content-addressed, the exact business context used in any run is provable after the fact. Underlying data quality is the source system's responsibility; the context that interprets it is the engine's.

02Types and classifications

Four knowledge types. Five classifications.

The knowledge type says what kind of claim a unit is. The classification says how it may change: a gate on every proposal to the graph.

Axiomatic

Definitions and rules that hold by construction: a chart of accounts, an entity hierarchy.

Empirical

Observed facts with a source: a balance, a rate, a measured outcome.

Prescribed

Policy and requirement: what must or must not happen, from a governing document.

Procedural

How work is done: steps, sequences, and conditions.

classification

gate on change

absolute: never auto-changes
regulated: needs a human
contextual: valid within a scope
tribal: conflict-checked
derived: updates itself

Linksrefinesdescribessame set

03Pipeline

Source. Extract. Store. Retrieve. Account.

01
Sources
Documents, connectors to systems of record, expert input, and past sessions.
02
Extract
Domain lenses turn source passages into Context Units: claim, evidence passage with source and location, knowledge type, classification, evidence grade, content hash. The model extracts; code validates every passage.
03
Store
Postgres as truth, Neo4j for typed links (refines, describes, same set), Qdrant for vectors. Sets scope knowledge to a topic, a business unit, or a use case.
04
Retrieve
Knowledge, Facts, and Search modes. A query agent plans the graph walk and repairs its own queries. No duplicate context across turns.
05
Account
Every unit is recorded as retrieved, injected into the prompt, or cited in the output. retrieved ⊇ injected ⊇ cited. Unknown citations are recorded separately.

04Retrieval

Three modes, one path for standing sets, mid-run search, and chat.

Knowledge mode walks the graph from a set anchor, fused and weighted by quality. Facts mode restricts the same units to those flagged as facts. Search mode runs semantic search over source passages. Standing sets preload before an agent's first turn; mid-run, the agent calls context search, pauses, retrieves, and resumes; chat uses the same path.

On the engine side, a query agent plans the graph walk and repairs its own queries, and a learning agent scores what was retrieved and used and flags units that never help for review. Nothing is deleted automatically.

Statusretrieval modes and engine-side agents

retrieval

knowledge · facts · search

KnowledgeSet-anchored graph walk, fused and weighted by quality.
FactsThe same units, restricted to those flagged as facts.
SearchSemantic search over source passages.

standing sets: preload before the first turn
mid-run: the agent calls context search, pauses, retrieves, resumes
chat: the same path
query agent: plans the graph walk · repairs its own queries
learning agent: scores what was retrieved and used · flags units that never help for review · deletes nothing

For engineers and architects4 min

05The accounting

Retrieved. Injected. Cited. Three numbers, every run.

For every run the platform records three tiers: what the context engine returned, what entered the prompt, and what the model named as support. The invariant is retrieved ⊇ injected ⊇ cited. A citation the model was never given is recorded separately rather than added. The accounting does not prove entailment; it gives an auditor a mechanically checkable trail, and because unit identities are content hashes, the trail names exact knowledge.

The same accounting feeds learning. Retrieved but never injected is evidence about ranking. Injected, cited, and wrong is evidence about the knowledge. Regulated and absolute units cannot be retired by usage signals.

Statusretrieved / injected / cited per runfeedback lanes from chat, frontend, and evals

run accounting · exec_01J9Q4M7Z3

retrieved

injected

cited

cited_unknown 0 · retrieved ⊇ injected ⊇ cited

cxu:1220a7ff…f843cxu:9b31e0d4…21aa

06Curation

Propose, gate, apply, observe. One claim over time.

A proposal comes from a document, a connector, an agent, or a person. The unit's classification gates it. Apply runs under compare-and-set: the new unit supersedes the old one and carries its hash; nothing is overwritten. Quality is then scored from use. The example follows one claim, an intercompany settlement term, through a policy change.

Propose01
a document, a connector, an agent, or a person
Gate02
by classification: absolute, regulated, contextual, tribal, derived
Apply03
compare-and-set: supersede, never overwrite; the new unit carries the old hash
Observe04
quality scored from use; never-used units flagged, not deleted

one claim over time · intercompany settlement

class regulated

2024-03cxu:8e02…19d1 · intercompany settlement net 60 · treasury-policy-2024.pdf · class regulated
2026-06-18proposal · treasury-policy-2026.pdf v4 · net 45 from 1 July 2026
2026-06-18gate · class regulated · waits for a person · no automatic change
2026-06-22r.okafor (treasury) approves · compare-and-set against 8e02…19d1
2026-06-22cxu:1220a7ff…f843 published · carries prior hash 8e02…19d1 · old unit preserved
2026-07quality scored from use · the learning agent flags units that never help for review

Statuscuration loopIllustrative

07Context Units

Content-addressed, evidence-backed, superseded, never overwritten.

Content-addressed

Identity is a content hash of claim and evidence. Identical knowledge produces the same identity across stores.

Evidence-backed

Every claim carries its evidence passage with source and location, and an evidence grade.

Typed and classified

A knowledge type says what kind of claim it is. A classification says how it may change.

Superseded, never overwritten

A change produces a new unit that carries the old one's hash. The old unit is preserved.

Composable

Units assemble into sets scoped by topic, business unit, or use case. Standing sets preload before an agent's first turn.

Graph-linked

Typed links (refines, describes, same set) let retrieval walk connected claims, not only similar text.

08Compared with retrieval over chunks

Typical RAG or agent memory, and cortIQ.

	Typical RAG or agent memory	cortIQ
Unit of knowledge	Text chunks sized for the embedding model	A claim with its evidence passage, source and location, type, classification, grade, and content-hash identity
Retrieval	Similarity search over one index	Knowledge, Facts, and Search modes; a query agent plans and repairs the graph walk
Provenance	A list of documents retrieved	Retrieved, injected, and cited recorded per unit, per run
Change	Re-embed the index	Propose, gate by classification, supersede under compare-and-set; the old unit is preserved
Improvement	None between re-indexes	Quality scored from use; units that never help are flagged for review, never deleted automatically

Go deeperWorkflows aren't enough: why knowledge is the moat8 min

09Shared across Apps

One graph for the tenant, filtered by scope on every read.

A monthly-close App runs extraction for one group's cycle. A week later an analyst in a fresh chat asks what changed; the copilot walks to the cycle and cites the exact units. A reconciliation workflow the next month consumes the same units as inputs. Nobody joins the original run. Authorization is applied before results are returned, so each caller sees only what its own scope allows, and an unauthorized unit is absent rather than redacted.

StatuscortIQ corebusiness knowledge plugins

10Questions

Frequently asked

What is cortIQ?

Pyrana's context engine, formerly the Context Engine. It turns documents, connectors, conversations, and system data into typed, classified, evidence-backed Context Units in a knowledge graph, retrieves them for agents in three modes, and accounts for what was retrieved, injected, and cited on every run.

What is a Context Unit (CxU)?

A claim, its evidence passage with source and location, a knowledge type (axiomatic, empirical, prescribed, procedural), a classification (absolute, regulated, contextual, tribal, derived), an evidence grade, a quality score, and an identity that is a content hash of claim and evidence. Units link in a graph: refines, describes, same set. When knowledge changes, a new unit supersedes the old one and carries its hash; the old unit is preserved.

How is this different from RAG?

RAG retrieves chunks sized for the embedding model, with no provenance and nothing that learns. cortIQ retrieves claims with evidence, keeps lineage under compare-and-set, and records per run which units the model was given and which it cited.

What context does an agent get?

Four layers. The agent configuration: a system prompt scoped to the use case. A library of files the agent discovers progressively. The context engine for company-specific knowledge. And, as a vision, business knowledge plugins (accounting principles, industry standards, epidemiology) curated as entity-agnostic context. Skills from public repositories are re-run through Agent Studio so they fit the platform's guardrails.

Who governs a change to knowledge?

Classification gates every change. Absolute never auto-changes. Regulated waits for a person. Tribal is conflict-checked. Derived updates itself. Proposals arrive from documents, connectors, agents, and people; feedback arrives from chat, frontend feedback lanes, and evals. Nothing is deleted automatically. Ownership at scale is a process Pyrana is still defining with clients.

How is sensitive knowledge protected?

Authorization is applied before results are returned, so an unauthorized unit is absent rather than redacted. Regulated and absolute units cannot be retired by usage signals; the learning agent flags, a person decides.

Does cortIQ work with any model?

Yes. Knowledge lives in the context layer, not in model weights. A model change does not rebuild the graph.

Company knowledge, typed, classified, and cited.

A demo extracts a client policy into Context Units and cites them in a run.

Book a demo Agentic Apps