Lab component — not in the Synedrix OS tech stack·Evaluated under LAB-001·Not required in any deployment mode

Not a Synedrix OS platform. LanceDB is not one of the twelve platforms and holds no OS layer. It appears here only as the embedded vector index of EverOS, which Tech Stack records as not adopted: LanceDB duplicates Qdrant, the single vector substrate for Layer 7. This page is lab reference material under LAB-001, evaluated in structural isolation from production — not a stack platform page, and not citable as authority by any canonical document.

What LanceDB is

LanceDB describes itself as a “developer-friendly OSS embedded retrieval library for multimodal AI,” built on the open Lance columnar data format. It is embedded rather than client-server by default: an application links the library directly (Python, TypeScript, or Rust) and reads/writes a table on local or object storage, with a separate REST-served mode available for multi-process access.

ConceptWhat it is
TableA Lance-format dataset — the unit LanceDB operates on, combining vector columns with arbitrary scalar columns in one row-and-column structure
Embedded modeThe default: no server process, the library reads/writes table files directly against local disk or an object store (S3, GCS, Azure Blob)
Hybrid searchVector similarity, full-text (BM25), and scalar filtering are queryable together against the same table, in a single query
VersioningLance’s columnar format keeps dataset versions, so a table can be time-travelled or rolled back without a separate snapshot mechanism

What LanceDB is normally used for

  • Embedded RAG applications — where a vector index needs to live inside the application process rather than behind a network call to a separate server
  • Multimodal retrieval — the format is built to hold vectors alongside images, video frames, and point-cloud data in the same table, not vectors alone
  • Local-first and edge deployments — no server to operate; a table is just files on a filesystem or object store

The common thread: LanceDB’s value proposition is specifically not having a separate database server. That is also precisely the property that makes it a poor fit for a platform whose organizing rule is one non-overlapping, centrally-operated store per responsibility — an embedded index by design does not participate in a shared, centrally-governed data plane the way a server-mode store does.

How Synedrix OS uses it

Synedrix OS does not use LanceDB in production. It is evaluated only inside LAB-001’s isolated environment, as the vector substrate that ships inside EverOS.

flowchart LR
  n1([LanceDB]):::term
  subgraph What it owns
    n2[Nothing in production]:::agent
  end
  n1 --> n2
  subgraph What it does not own
    n3[Semantic retrieval and similarity search]:::warn
    n4[Qdrant]:::term
    n3 --> n4
    n5[Canonical relational truth]:::warn
    n6[Supabase PG17]:::term
    n5 --> n6
    n7[Explicit relationships]:::warn
    n8[Neo4j]:::term
    n7 --> n8
  end
  n1 --> n3
  subgraph Why LanceDB is not in the stack
    n9[Substrate duplication not capability]:::term
    n10[Inseparable from the runtime it ships in]:::term
    n11[Worth evaluating precisely because it is not adopted]:::term
  end
  n1 --> n9

What it owns

Nothing in production. Inside LAB-001’s isolation boundary, a lab-local LanceDB instance backs EverOS’s own vector index — reachable only from lab infrastructure, never from a registered agent, and holding no production tenant data.

What it does not own

ConcernOwner
Semantic retrieval and similarity search — the single vector substrate for Layer 7Qdrant (Layer 7)
Canonical relational truth — business entities, workflow state, approvals, audit trailsSupabase PG17 (Layer 7)
Explicit relationships — workflow and outcome lineage, agent influence networksNeo4j (Layer 7)
Anything a registered agent may recall as memoryThe governed memory fabric (Memory Fabric Spec), never a lab component directly

Why LanceDB is not in the stack

  • Substrate duplication, not capability. ADR-OS-002 §5.4 states the deciding axis plainly: “substrate duplication, not capability” — the same test that separated LangGraph from LangSmith on licensing terms, applied here to storage. LanceDB is a capable embedded vector store; the OS already has a vector store, and its organizing rule is one per responsibility.
  • Inseparable from the runtime it ships in. ADR-OS-002 §5.4: “EverOS cannot be reduced the same way [as NAMS] — its value is inseparable from LanceDB and SQLite.” NAMS’s own vector index could be scoped away and the library kept; EverOS’s cannot without losing the feature that makes EverOS worth evaluating in the first place, which is why EverOS’s model was adopted and its runtime was not.
  • Worth evaluating precisely because it is not adopted. LAB-001 exists to test, in isolation, whether that rejection’s practical cost (operational burden of a second vector store) is as high as the argument assumes — a question worth answering with evidence rather than assumption, without reopening the decision to get there.

Deployment and licensing

Self-hosting. LanceDB’s embedded mode requires no server deployment at all — the library is a dependency of the application process. A separate server mode exists for concurrent multi-process access. Both modes are self-hostable; neither is relevant to a Synedrix OS deployment mode, since LanceDB is not part of the stack.

Licence — verified. LanceDB is licensed under the Apache License, Version 2.0 (“Apache License, Version 2.0, January 2004”), confirmed against LICENSE in the official lancedb/lancedb repository. This matches the licence position ADR-OS-002 already recorded and requires no correction.

Official documentation

  • LAB-001 — EverOS Lab Charter — The isolation boundary and evaluation questions this page’s evaluation runs under.
  • ADR-OS-002 — Memory Fabric Control Plane — §3.1 and §5.4, the decision that rejected LanceDB as part of the canonical stack.
  • ADR-OS-004 — Lab Lane — The governance mechanism this evaluation runs under.
  • Qdrant — The single vector substrate LanceDB would duplicate.
  • Memory Fabric Spec — §10, the invariant that no second vector store enters Layer 7.

Synedrix OS · Platform reference · Lab · LanceDB