@volter/twin-pinecone

Legacy connector helpers: this package still has callable helpers using the retired v1 syncPull API. Those paths require migration before use on the current kernel; older helper descriptions below do not establish current compatibility. Check the generated index for protocol standing and use the shared model for current state semantics.

Local Pinecone twin for the vector-database surface: index/collection management on the control plane (api.pinecone.io), vector upsert/fetch/update/delete, a REAL brute-force top-K similarity query (cosine / dotproduct / euclidean — exact, not approximate) gated by a REAL Mongo-style metadata filter evaluator, namespaces, describe_index_stats, and host-split control/data-plane routing — built on the shared @volter/world-core kernel. API-first vendor (no product UI to mirror — the real surface agents integrate with is the REST API / @pinecone-database/pinecone SDK, not the Pinecone console).

bun packages/twin/pinecone/src/cli.ts

Coverage

This is a v1 slice of the Pinecone API, not the full surface. Modeled done: index management (create/list/describe/configure/delete, a deterministic minted per-index data-plane host), collection management (create/list/describe/delete), vector upsert/fetch/update/delete (by-ids, delete-all-in-namespace, delete-by-filter), vector list, query with all three real similarity metrics (cosine, dotproduct, euclidean — hand-computed expected scores/ordering in the test suite, including a crafted vector set where cosine and dotproduct rank neighbors differently and a filter that excludes the otherwise-nearest match), namespace isolation, describe_index_stats, and the host-split control/data-plane router.

Left as todo (honest gaps, not yet modeled): sparse/hybrid search + reranking, $exists / deep-nested metadata filters, pod-based index configuration (replicas/pod-types/serverless-vs-pod choice), deletion-protection enforcement, collection-from-index full snapshot/restore semantics, bulk import (start/status/list/cancel), backups/restore, integrated inference (embed/rerank) and the integrated-embedding surface (createIndexForModel, /records/{namespace}/upsert|search), namespace list/delete as their own endpoints, real cursor pagination on vectors.list, the full gRPC-status error-details envelope, the control-plane error-shape's exact wire format, API-key 401 auth parity, real (non-zero) indexFullness, connector push, and a fixture-seed helper.

See src/pinecone-capabilities.ts for the full manifest (the real vendor surface is the denominator — coverage is honest and partial until the twin reaches it).

A note on the query engine's honesty

Unlike this repo's generative twins (Anthropic/OpenAI/Replicate/fal), Pinecone's core operation — nearest-neighbor search over stored vectors — is genuinely computable offline: given the stored values, exact cosine/dotproduct/euclidean top-K is deterministic, correct math, not a placeholder. src/pinecone-similarity.ts and src/pinecone-filter.ts are net-new modules (not produced by cloning the replicate skeleton this pack started from) implementing that real compute — see their file headers for the exact grounded semantics, including the one doc- UNVERIFIED item (euclidean's exact wire-level score/sign) called out inline.

No webhooks, no signed requests

Pinecone's v1 REST surface has no webhook/signed-request surface (unlike Replicate/fal), so this pack has no webhook module and no pure-crypto mutation-test allow entries.