Context Farm

Worked example / synthetic data

The rule. The exception. The source.

Choose a dispatch question to see a prepared answer grounded in the prototype's existing sample objects. This is an illustrative rules walkthrough, not a live AI response or production dispatch advice.

The Problem

Most small teams do not have a documentation shortage. They have a knowledge concentration problem. One person knows the real workflow, the exceptions, the caveats, and which source actually wins when documents conflict.

That breaks teams in predictable ways: repeated interruptions, slow onboarding, inconsistent execution, and AI agents that can read the handbook but still make bad decisions. Context Farm exists to turn that concentrated operational knowledge into reusable context.

Boundary Labs hit the same wall internally. Retrieval was finding relevant documents, but the agent or operator still needed the unstated rule, the exception, or the ranking between sources. The chunks were there. The operational truth was not.

Deployment boundaries

Local deployment is a design goal. Storage location and inference routing are separate choices: a local knowledge store does not itself guarantee that no data reaches an external model. Agree on both boundaries before using sensitive material.

How It Works

Context Farm is now being developed as a layered system: ingest messy operational material, compile it into readable knowledge artifacts, extract typed operational objects, and serve grounded retrieval and briefings to humans and agents.

CURRENT DIRECTION
─────────────────────────────────────────────────────────────

  raw input          │  PDF, URL, text paste, transcripts, operator seed
        ↓            │
  ingestion          │  normalize source, preserve provenance, assign authority
        ↓            │
  compile layer      │  build readable linked artifacts from messy material
        ↓            │
  structured layer   │  extract facts, procedures, constraints, exceptions,
                     │  decisions, and source-linked evidence
        ↓            │
  governance         │  review high-impact items, track authority, flag conflicts
        ↓            │
  serving layer      │  search, ask, brief, and agent retrieval

─────────────────────────────────────────────────────────────
STORES

  wiki / article layer   │  human-readable audit trail and fallback retrieval
  SQLite                 │  structured operational objects
  ChromaDB               │  semantic recall over compiled knowledge

Domain seeding

The key differentiator is still domain seeding. Before full ingestion, the operator describes the domain in plain English: what matters, what entities exist, what rules apply, what exceptions are common, and which sources outrank others. That seed guides subsequent extraction and review.

Why the article layer stays

The wiki-style compile layer is not dead weight. In the current design it serves four jobs: human-readable audit trail, intermediate normalization before structured extraction, fallback retrieval while extraction is incomplete, and a debugging surface when the structured layer gets something wrong.

Inputs & Current Scope

Current State

Two things are true at once. First, the underlying ingestion and knowledge-compilation pipeline has been running internally across finance, research, and infrastructure domains. Second, the product-shaped Context Farm work is now being tightened around a small-team operational-memory use case with an explicit manual demo domain before broader extraction automation.

Current Gaps

Context Farm has real forward motion, but the hardest parts are not being glossed over. These are the actual gaps now driving the build.

Extraction quality is the product

The SQLite schema and manual demo path are in place. The hard part is the extraction loop that turns raw material into the right object type with enough provenance and low enough review burden to be trusted. Constraint versus exception versus procedure is not a trivial distinction. This is the current center of gravity.

Governance without turning into a chore

Review is necessary, but a small-team operator cannot spend their day approving rows. The system has to batch high-impact review, auto-accept low-risk items where possible, and surface contradictions only where the review load is justified.

Schema drift

Domain seeds are useful, but domains change. New entities appear, old exceptions stop applying, and source hierarchies get messier over time. Detecting when the model of the domain is stale is still open work.

Cross-domain and freshness handling

The internal foundation already spans finance, research, and infrastructure, but the product path is deliberately narrowing before it widens again. Cross-domain retrieval and staleness tracking both matter, but they come after the single-domain operational-memory loop is credible.

Contact & Partnerships