Real local embeddings and remote retrieval — 2026-09-29

This is development qualification for #72, before the final RC artifact. It uses actual learned embeddings and native subscription accounts. No provider substitute, fixed vector, canned model response, private library or private mailbox was used.

Model and documents#

The optional runtime is Transformers.js 4.3.0, with CPU ONNX inference using Xenova/all-MiniLM-L6-v2, revision 751bff37182d3f1213fa05d7196b954e230abad9, q8, mean pooling, 384 dimensions and a 256-token input bound. Authentication is none; rc-local-embeddings is a local deployment identity, not an invented provider login. Real token counts use adapter-report through the existing usage-envelope contract. No API bill is inferred from local computation.

• The complete 19-page Lewis et al. RAG paper was extracted by installed Poppler, with no omitted or OCR pages, into 191 passages. The original PDF SHA-256 is recorded in the JSON receipt.
• Repository release-gate Markdown yielded eight passages with section/line provenance. A realistic two-message release-handoff thread yielded two passages retaining the thread and message identities.
• Prepared documents used a maximum of 512 UTF-8 bytes per passage. Model input limits were enforced by the actual tokenizer rather than an estimated count.

The complete local sequence passed semantic similarity against an unrelated passage, cancellation and subsequent execution, rejection of oversized token input, missing pinned-model rejection, unchanged-document indexing without another embedding call, namespace/selected-source isolation, immutable revisions, stale-evidence rejection, explicit reindexing, deletion and incompatible-index rejection. Closing/reopening SQLite returned identical evidence. A separate real check filled the four-call queue, rejected a fifth, cancelled a queued call and completed the remaining work. Explicit preparation in a fresh cache downloaded the pinned real model successfully. Normal calls used offline loading.

Remote protocol and generation#

A separate Node 24.16.0 host on Prometheus indexed the prepared passages using the same real local model. The original PDF extraction ran on the application computer; remote calls indexed its actual prepared chunks. The host bound only loopback 17437, reached through verified OpenSSH. It used its own connection store, database, grants and usage files. Other application/validation hosts and native account profiles were preserved. The reviewed six-hour grant covered only the two selected native providers and the release-library corpus's search/index/delete operations; it granted no tools, management or jobs.

TypeScript, Python, Go and Rust clients independently returned identical four-hit search results, including model identity, original source revision, page/line locations, ingestion manifest and passage digest. TypeScript used the recorded development archive; Python used a freshly installed wheel; Go used published alpha.6; Rust used development source after adding serialization for its typed retrieval results. This discovered and fixed the missing Rust Serialize implementation. These are protocol checks, not a claim that final RC client artifacts have already been published or accepted. Restarting the real remote host preserved the same scoped connection and exact SQLite search results.

Generation explicitly selected the already signed-in Codex CLI 0.157.0, gpt-6-luna, medium reasoning, and Claude Code 2.1.282, claude-haiku-4-5-20251001. Both reported Pro subscriptions. Catalog refresh reported nine Codex and fourteen Claude entries; only the selected models ran.

The first failure is retained. A separate request supplied the actual paper question as the retrieval query, keeping writing instructions out of similarity search; it did not switch account, model or billing mode. No automatic generation retry ran. The final paper answer distinguishes document-level versus token-level marginalization and labels its attribution implication as an inference. The email draft retains pending PyPI approval and no-private-mail/no-send commitments. Both remained drafts; no application acceptance or external email was performed.

The SDK now exposes opt-in exact citation-ID validation in all four languages. Rechecking the original real responses accepted Claude's IDs and rejected Codex's typo with INVALID_CITATION, without another model call. This verifies reference identity, not factual support. Applications still review claims and quotes and recheck current source permissions before accepting an artifact.

Codex's second run reported 4,864 cached input tokens. Claude's reported USD 0.02163 API-equivalent amount is not an actual subscription charge. Reported output usage need not equal visible answer length. No default deadline, inactivity timeout or provider fallback was introduced.

Evidence and limits#

The sanitized JSON receipt records prompts, model/index identities, run IDs, counts and checksums. Full responses, retrieved paper passages, prepared chunks and private profiles remain in private local validation state. Source/batch limits are bounded exact cosine search, not a large-corpus ANN performance claim. Only this English model and Linux route are qualified here; other models/platforms and OCR need their own checks.

Local typechecking, 27 pure SDK tests, installed Python/Rust/Go contract checks, fresh npm installation without the optional model runtime, and the documentation build passed separately. Integrated security review and final RC artifact/app acceptance remain release gates.

The implementation commit a8c53a6a23e3d961aa9659d9ac7c51f795f1b34d passed all seven checks in Prometheus run 36562335687, attempt 2. The first attempt stopped four jobs at the private disk-space guard; SDK-owned retired temporary archives and unused CI images were cleaned before rerunning. Successful jobs were retained, failed jobs reran successfully. No credentials or model requests were added to CI. Final RC artifact acceptance remains #71.