The axes that actually matter.
Side by side.
You're comparing us against Glean, Onyx, Bedrock KB, Azure AI Search, Gemini Enterprise, or a build-it-on-MCP stack. The vendor decks all read the same. This page is the part they leave out — the architectural decisions that change what you can ship on top.
The market offers two deals.
Both come up short.
Every team trying to give AI access to "all our stuff" today ends up in one of two boxes. inseam is not a third box — it's a different layer underneath.
- Retrieval is whatever the SaaS search API exposes — N shallow per-source endpoints.
- Connector catalog is gated by the agent vendor's engineering capacity.
- Permissions inherit the SaaS's ACLs — workspace-binary, not party-scoped.
- No cross-source dedup, no normalized identity, no offline anything.
- A second copy of every piece of your data, chunked and stored on the vendor's hardware.
- Proprietary, opaque ranking. You can't introspect why a result was chosen.
- Workspace-shaped permissions; safe multi-party retrieval is bespoke per customer.
- Cloud-first by default. Self-host is "available" but server-shaped, not device-shaped.
Eight comparisons.
Each one a claim that has to hold up.
If a competitor scores well on all eight, they're a direct threat. The full matrix
and the per-vendor breakdown lives in
arch/meta/market/competitors.md; this is the short read.
Every standard RAG pipeline stores chunk text alongside the embedding. The vector DB becomes a second copy of the corpus. That copy is what forces the vendor to be a cloud — it's too big to live anywhere else, and the licensing story around copied content is a problem most vendors hand-wave through. inseam doesn't copy. The body is fetched live, on demand, through the Connection that owns the credential.
Not a fan-out of N per-source searches glued together by an LLM at query time. Cheaper at inference, better recall, introspectable. The hierarchical summary tree pairs natively with small on-device models — the same retrieval shape the research community is converging on (A-RAG, RAPTOR, progressive disclosure).
External identities — customers, clients, counterparties — inferred from data envelopes get scoped read paths into the index without per-customer plumbing. Workspace-binary ACLs are a different shape. This is the axis where Glean, Onyx, Bedrock, and the agent-side stack all share the same gap.
Two adapters today: SQLite on a device, D1 in a Worker. No local-vs-cloud code fork. Cross-device pooling is first-class — a cloud node can index documents it has never received the body of, with fetch hopping to whichever node currently holds the live Connection. No competitor in either category models this.
Vendor-side connector catalogs are bounded by headcount. inseam's is bounded by how many users have an agent and ten minutes. The contract is small on purpose — small enough for an agent to fill in correctly, with framework-enforced fixtures and tests as the quality bar.
No proprietary index, no enterprise edition, no investor to please with a pivot. Forkable. The procurement story stops being "is this vendor still here in three years" and starts being "do we like the code."
Closest single competitor architecturally: Onyx — open-source, self-hostable, RAG-grade. Differs on the four pillars that define inseam: copies the corpus, server-shaped not device-shaped, no cross-device federation, workspace-level permissions.
Run a POC against the axes that matter.
The full design docs live under /arch. The competitor matrix and
per-vendor breakdown live under arch/meta/market/. If a comparison axis
isn't covered, open an issue — we'd rather have the conversation in the open than
in a sales call.