Update Readme

[ragnorc]

No description provided.

[ragnorc]

Recommendation: approve. I found no blocking defect. This is a COMMENT review because the authenticated account is also the PR author.

This PR changes how the README introduces OmniGraph. It names object storage, typed schemas, branches, and the three retrieval modes directly. It also removes a duplicate crate list. Readers get a shorter introduction, while the architecture guide remains the owner of component details. The PR changes no runtime code or storage behavior.

The new claims match the implementation:

• The workspace pins Lance 11.0.0 with AWS and Azure support. The pinned Lance provider registry supplies local, S3, and Azure access. OmniGraph's deployment guide retains the Azure qualification limits.
• The query lowering handles nearest-vector and BM25 ranking, plus graph expansion. These are existing operations within one execution path.
• “Typed ontology” describes the declared node, edge, and property schema. “Open data format” describes the Lance substrate. Neither phrase adds an inference engine or permits arbitrary external writers. OmniGraph still selects coherent table versions through its graph catalog, as the architecture guide explains.

The tradeoff is less introductory context in exchange for more concrete feature names. The agent use cases remain immediately below the heading. Those workloads combine retrieval with isolated branch changes and review before merge. This edit makes no new latency, scale, or concurrent-writer guarantee. It also removes the nearby crate map, so a direct architecture link would help contributors find that information.

Liability decreases slightly. The net reduction of 12 lines removes a second component inventory that already omits the planner, core, and catalog crates. The benefit comes from one maintained source of truth, not the line count alone. No API, state, invariant, or abstraction needs maintenance. Five similar edits should replace repeated details with links to their owners.

Validation on exact head 98481e9174f659298f314403fa8161f4dee51789, in a clean isolated checkout:

• Local documentation check passed for 203 Markdown files. AGENTS links and diff whitespace checks passed. No repository Markdown link targets the removed heading.
• I checked the relevant OmniGraph code and test owners, full upstream overview/search/versioning/storage guides, and pinned Lance implementations.
• CI for this head reports successful documentation links and storage-upgrade checks. The separate GQT run also passed. The workspace test job was skipped.

I did not run Rust suites, cloud tests, or performance tests locally for this documentation-only change. Source inspection supports the feature descriptions, not a new performance claim. The inline suggestion is optional.

[ragnorc]

Recommendation: approve with two optional wording corrections. I found no blocking defect. This is a COMMENT review because the authenticated account is also the PR author.

Reviewed exact head 7de8ce5e7cf652db46d6011fb9448e3a700e1324 in a clean isolated checkout. This follow-up covers the six README commits since 98481e91.

The new introduction explains why agents would share OmniGraph: they retrieve context, store durable changes, and coordinate through a common typed model. It puts four short feature descriptions before the use cases and moves the existing capabilities table below them. Readers get the workload before the deployment detail. The PR changes no execution, storage, or API behavior.

The underlying mechanisms support that direction. Query lowering combines vector and BM25 ranking with graph operators. Branches isolate proposed changes, while the existing merge path publishes the result. Typed schemas enforce the declared data model. Branch review is a supported workflow, not a mandatory approval gate: the mutation request still defaults to main. Multiple agents also do not imply multiple independent writer processes. The documented branch-control boundary still applies.

Two claims go beyond those mechanisms. Combined retrieval does not establish high precision and recall for an arbitrary corpus. Object storage does not establish the lowest total cost. The inline comments suggest concrete replacements. These are documentation improvements, not runtime defects.

The tradeoffs are explicit:

• Short feature descriptions make the intended use easier to understand. They omit limits that the search, branching, and deployment guides own. Links to those guides would keep the introduction concise and verifiable.
• ANN search trades recall against latency and object-store I/O. OmniGraph passes the plan's probe cap to Lance and widens short searches. Finding enough candidates does not prove exact ranking or application relevance. See the scan adapter and pinned Lance implementation.
• Shared durable state can simplify clients. The server still owns query execution, policy, and publication. Storage requests, compute, index maintenance, and retained branch history still contribute to cost.

Runtime liability is unchanged. Documentation liability increases slightly: the four bullets repeat the capability table and add two unqualified outcome claims. This follow-up adds 19 lines and removes 12. The earlier removal of the duplicate crate inventory still reduces maintenance. Five similar edits should preserve one detailed guide per feature and link to it, rather than repeat stronger promises.

Validation: local documentation checks passed for 203 Markdown files. AGENTS links and diff whitespace checks also passed. I checked the relevant code, existing cost-test scope, and the pinned Lance 11 implementation. The complete upstream guides and repository guides from the earlier review remain applicable. Only README changed. CI for this head reports successful documentation links and server tests. GQT also passed. Storage Upgrade Compatibility is still running, and the workspace test job was skipped.

I did not run Rust, cloud, relevance, or cost benchmarks locally for this prose-only change. Existing I/O-count guards do not prove the new comparative cost claim. The checkout remains clean.

[ragnorc]

Recommendation: approve the latest README changes. I found no new blocking defect at a8ac5ccd1a37326c87ed89460540a72413546d00. This is a COMMENT review because the authenticated account is the PR author.

This follow-up covers the changes since 7de8ce5e. The README now links to an agent-readable documentation index, separates the short introduction from the feature table, and removes a duplicate agent-fleet row. It also shortens the configuration summary and fixes the heading typo. Readers get a clearer route from the intended workload to detailed instructions. Only README.md changes.

The intended workload remains many agents reading context and proposing changes to shared, durable graph state. The new statement that agents and UIs “can become” stateless consumers describes a client design option. It does not promise a stateless database server or independent concurrent writer processes. The writer topology still requires one mutation-capable writer process per cluster.

I checked the shorter configuration claim against the actual configuration types. They declare graphs, schema paths, stored queries, embedding providers, and policy bindings. The later deployment section still explains how those declarations become a running service.

The tradeoff is less detail in the feature table, with readers following the guides for operational limits. The new llms.txt link improves discovery but depends on the hosted documentation remaining available and current. It does not copy another detailed manual into this repository.

Runtime liability is unchanged: no new API, state, primitive, or storage behavior. Documentation liability decreases slightly because the duplicate row is gone. The full PR adds 23 lines and removes 26. The reduction helps, but the stronger evidence is less repeated content. Five similar edits should keep detailed behavior in its owning guide and use links from the introduction.

The earlier optional comments about retrieval quality and lowest cost still apply. Those claims remain unchanged. The inspected OmniGraph scan path and pinned Lance 11.0.0 implementation support approximate retrieval, not a general relevance or cost guarantee. I did not repeat the inline comments.

Validation:

• Local documentation checks passed for 203 Markdown files. AGENTS links and diff whitespace checks passed.
• The new documentation index returned HTTP 200 with plain-text content. All 22 linked documentation pages also returned HTTP 200.
• The checkout is clean. Repository instructions, runtime code, tests, and the Lance pin are unchanged from the previous reviewed head. I reused the earlier full upstream documentation review and rechecked the relevant implementation.

I did not run Rust tests or performance benchmarks for this prose-only change. The link checks establish current availability, not continued availability or full documentation accuracy.