Agentic Data

The problem

Enterprise knowledge is dying. Every organization generates a continuous stream of institutional knowledge — pull requests, incident timelines, RCA documents, architecture decisions, customer escalations, deployment logs. These artifacts are scattered across GitHub, Jira, Slack, Confluence, PagerDuty, ServiceNow, and dozens more systems.

The connections between them — this PR caused that incident, which led to this RCA, which changed this decision, which was later reverted — live only in people's heads. When those people leave, the connections disappear. The organization retains its documents but loses its understanding.

This is not a search problem. You can find the documents. What you can't find is the meaning between them.

Not RAG

The retrieval layer we have today — embed, top-k, stuff into prompt — is fundamentally insufficient for enterprise knowledge.

• Documents are not the unit of knowledge. Events are. A PR, an incident, a decision — these are episodes that happened at a specific time, involved specific actors, and had specific consequences.
• Similarity is not relevance. When you ask "why did the payment service go down?", the answer is a causal chain through time, not the 10 most semantically similar paragraphs.
• Retrieval is not a single call. It is an iterative control loop that plans what evidence it needs, searches, evaluates what's missing, expands through graph traversal, and stops when sufficient — or admits that it isn't.

Three interlocking ideas

35 typed edges

Each edge is directional, carries a confidence score (0.0–1.0), provenance method (deterministic rule, LLM extraction, human verified), evidence spans pointing back to source text, and validity windows.

Agentic retrieval loop

Instead of embed(query) → top-k → stuff into prompt, the retrieval planner runs an iterative control loop that hunts for a sufficient evidence graph.

The loop is budget-aware (max retrieval calls, graph hops, latency, model calls) and stops on three conditions: sufficiency threshold reached, budget exhausted, or diminishing returns detected.

Intent-driven strategies

The retrieval planner selects a strategy based on the query's intent, determining which edge types to prioritize during graph expansion.

What is live now

Agentic Data is operational. The following infrastructure is built and tested.

Architecture

Four data stores, each for what it's best at. FastAPI + SQLAlchemy 2 (async) ties the layers together with JWT auth, RBAC, rate limiting, and Docker Compose for local and staged deployments.

Graph query API

Beyond flat search, the graph API provides structured traversals for the questions that actually matter in enterprise operations.

Truth maintenance

A knowledge system that only accumulates and never revises will inevitably hallucinate with confidence. The canonical memory layer is self-correcting.

• Supersession chains — when a new note replaces an old one, the chain is preserved so you can trace how understanding evolved
• Deprecation markers — disputed or outdated notes are flagged, not deleted, maintaining the full audit trail
• Contradiction detection — new evidence is checked against existing canonical notes for conflicts
• Reconsolidation loops — detected contradictions trigger review workflows: suggest update, approve/reject, verify
• Validity windows — notes and COs carry explicit valid_from/valid_to bounds; expired items are automatically caught by maintenance jobs

Connectors

Source system integrations that fetch items updated since a timestamp and ingest them through the full pipeline (persist → extract edges → index → graph sync).

Design philosophy

• Relevance is not similarity. Two documents can be semantically identical and one can be wrong (superseded, reverted). Relevance is defined by causality, temporal validity, authority, and organizational state.
• Permissions are not a bolt-on. ACL filtering happens before any model sees any content. Graph traversals cannot leak restricted nodes. Every piece of evidence carries provenance.
• Don't just retrieve more. If answer quality degrades as you add context, the system collapses under scale. Aggressive reranking, evidence budgets, and sufficiency-based stop conditions are architectural requirements.
• The graph is the product. The context graph is the durable asset that compounds over time. Every ingested document makes it richer. Every query validates or challenges existing connections. This is the flywheel.

Roadmap

The graph is not an implementation detail.

It is the durable asset that compounds over time.

Every ingested document makes it richer. Every query validates or challenges existing connections. Every new edge makes future queries more precise. This is the flywheel.