Hierarchy Plus Time: The CHORUS Approach to Graph RAG

But CHORUS isn’t just a smarter retriever. It’s a system: temporal addressing, curated ingestion with justifications, agent teams, and opinionated security/on-prem deployment. Same melody, very different arrangement.

Most “HiRAG/GraphRAG-like” approaches make three sound bets:

Structure beats soup. You cluster/segment knowledge, generate layered summaries, and route queries top-down instead of brute-forcing a flat index.
Graphs matter. Entities, relations, and communities give the LLM handles for multi-hop reasoning.
Coarse → fine retrieval. Start with abstractions to decide where to look; only then pull the receipts.

CHORUS agrees with all of that—and uses similar patterns when it makes sense. We’re not here to reinvent gravity.

Where Hierarchy Alone Falls Short (and where CHORUS compensates)

1) Time is part of the address, not an afterthought

Most hierarchical RAG stacks don’t model time beyond a “latest wins” snapshot. CHORUS does.

UCXL temporal navigation lets you move backward (~~/) and forward (^^/) across versions as part of the address itself.
The hierarchy isn’t just “what’s relevant now”; it’s how relevance evolved, which matters when you’re auditing a decision, migrating policy, or comparing strategies across quarters.

Net effect: You don’t just answer what—you can answer when it became true and why it changed.

2) Curation > clustering (on purpose)

Unsupervised clustering is useful. It’s also blunt. CHORUS layers SLURP (a curated ingestion pipeline) on top:

Curated roll-ups: Summaries and bundles aren’t just LLM blur. They’re vetted artifacts with justifications—citations tied back to sources and prior decisions.
Policy-aware shaping: Sensitive content gets shaped, redacted, or split before it enters retrieval. That’s not “secret sauce,” that’s responsible engineering.

Net effect: Your hierarchy is explainable and enforceable, not just plausible.

3) Evidence with an address

CHORUS stores context as encrypted Markdown in a hierarchical, addressable space (HCFS/UCXL). That means:

Every summary, edge, claim, or decision has a stable, dereferenceable address.
Teams can link to evidence rather than re-summarizing it—so downstream agents (or humans) can re-check the original in context.

Net effect: Less “trust me, bro.” More “here’s the exact artifact.”

4) Teams of agents, not loner prompts

Hierarchical RAG often focuses on a single model’s retrieval. CHORUS is intentionally multi-agent:

WHOOSH/BZZZ orchestration (roles, team channels, consensus to submit) treats complex tasks like software projects, not single prompts.
Agents use the same address space, so their hand-offs maintain context, lineage, and time.

Net effect: Complex work gets done collaboratively, with durable context—not one-shot magic.

5) Opinionated security and on-prem reality

Most papers are cloud-agnostic. CHORUS assumes the opposite:

On-prem friendly by design; encrypted, need-to-know context; p2p/DHT transport so there’s no single choke point.
Clean lines for licensing and audit (who accessed what, when, and why) without shipping your crown jewels to someone else’s GPU.

Net effect: You get modern retrieval and reasoning without bleeding IP into the cloud.

6) The economics of many small brains

Yes, large models are powerful. They’re also expensive and often unnecessary.

CHORUS routes work across networks of smaller, specialized models where it’s sane to do so, reserving heavy models for the steps that truly need them.
The orchestration plus addressable evidence makes this auditable (and kill-switchable) when costs spike.

Net effect: Better answers, fewer unpredictable bills, clearer trade-offs.

A Concrete Contrast (minus the secret herbs)

A typical hierarchical RAG flow

Build clusters → write layered summaries → answer queries by drilling down.
Works well for “what is X?” and many multi-hop facts.
But it treats time, governance, and provenance as side quests.

The CHORUS way

Ingest into UCXL with curated summaries and justifications (SLURP).
Build hierarchies per epoch so addresses can jump across time (~~/, ^^/).
Agents plan work, cite addresses (not ephemeral chunks), and reach consensus before shipping artifacts.
Security, audit, and on-prem constraints are table-stakes, not plugins.

No secret formulas there—just system choices that close the gap between “great demo” and “ship it in a real company.”

Where We Intentionally Align

We’re not contrarian for sport. We actively use hierarchical retrieval:

Top-down routing to cut token waste.
Community-local hops to stay inside the right neighborhood.
Entity-exact pulls when the job needs the raw line.

CHORUS treats this as a pluggable retriever behind UCXL/RUSTLE—valuable, but not the whole story.

What You Get That “Hierarchy-Only” Won’t Give You

Version-aware answers: “What changed since 2023—and who approved it?”
Reproducibility: The same UCXL address tomorrow yields the same artifact (or a clearly newer one you can diff).
Compliance and explainability: Summaries come with justifications; agents cite addresses you can actually open.
Operational sanity: Multi-agent workflows, not prompt spaghetti.
Security posture: On-prem, encrypted, need-to-know. Period.

Final Take

If you already like hierarchical/graph RAG, good—you should. CHORUS embraces that foundation. But we’re building for accountable, time-aware, multi-agent work under real security and cost constraints. That requires more than a clever index.