keble-idea 0.8.0

Provider-neutral idea extraction, concept graph, generation, and search contracts.

Keble Idea

keble-idea owns the provider-neutral contracts and application behavior for public content snapshots, standout ideas, semantic concepts, relationships, immutable generations, durable work ledgers, and concept-first hybrid search.

It deliberately does not import FastAPI, Celery, MongoDB, Redis, Qdrant, Neo4j, raw providers, or keble-data-infra. The host application implements the protocols under keble_idea.application.ports and converts raw source responses into idea-owned schemas at its outer boundary.

Version 0.5.0 changes immutable model execution from one physical deployment to one Infisical-catalog logical model. Every model-backed envelope, receipt, and accepted relationship freezes its purpose, ProcessingTier, selection rule, logical model/provider keys, model revision, catalog fingerprint, and generation. Every real provider request separately records a secret-free ProviderAttemptBinding with endpoint-pool revision, endpoint, deployment, region, and billing mode.

Version 0.6.0 aligns durable failure evidence with the universal raw contract: a physical HTTP rejection may truthfully have complete zero billable usage (for example reviewed authentication, permission, payment, invalid-request, or rate-limit rejection). A zero-attempt local gate still cannot carry usage, and USAGE_ONLY evidence remains financially incomplete.

Version 0.6.1 makes physical-attempt budgeting explicit on the immutable failure snapshot. A zero-attempt PLATFORM_RUNTIME observation—such as waiting for an already-paid completion projection—does not consume another provider attempt. Provider/adapter failures and missing evidence remain conservative and consume the budget. The wall-clock deadline still terminates local polling.

Version 0.6.2 completes the REVIEW2 replay boundary. Publication repositories return one canonical hydrated result for both first publication and receipt replay, so a redelivered paid extraction retains every idea, concept, and posting identity needed for downstream projections. Durable upstream failure evidence also keeps the authoritative classification origin and provider mapping revision, while separately metered identity correction remains owned by ExtractionService and is included exactly once in terminal receipt usage.

Version 0.7.0 introduces one evidence-backed semantic corpus for seller and shopping discovery. COMMON and UNCOMMON ideas and facets pass the same quality gates, persist normal concepts and postings, and share one vector/graph projection. SearchCorpusMode.DISTINCTIVE_ONLY is the seller-safe default; ALL_DETERMINED explicitly admits common purchase attributes for shopping search. UNDETERMINED content retains compact assessment, receipt, usage, and lineage facts but produces no searchable semantics. Content summary is derived as COMMON, UNCOMMON, MIXED, or UNDETERMINED only after publication gates.

Version 0.8.0 combines the separate provider-neutral content catalog with the durable restart-recovery contracts in one package candidate. Catalog projection uses PROJECT_CONTENT_SEARCH plus CONTENT_EMBEDDING; recovery extends the same canonical envelope with lease-renewal, expired-attempt planning, and finite sweep reports. It adds no parallel job state machine and remains unreleased until the coordinated Platform candidate and owner relevance gate are complete.

The content-catalog contract adds a separate provider-neutral catalog-search surface without changing semantic idea discovery. Both extraction result variants now return up to 48 evidence-grounded human term units inside the existing metered model call. ContentSearchProfileBuilder merges those proposals with typed structured content facts by canonical NFKC/casefold identity, preserves multi-word terms, caps only after merge priority, and builds one deterministic token-measured embedding input for every classification, including UNDETERMINED.

Content-profile publication is receipt-last: a terminal result is complete only when its deterministic profile row also exists, and replay returns that exact profile key. Host adapters implement the Idea-owned projection, ordered Mongo hydration, Redis session, cursor-codec, and runtime-resolution protocols; ContentSearchService composes them into the one new-search/cursor workflow. It pre-generates the search-session identity used for semantic cost lineage, probes one row beyond the configured result window, freezes only unique compact frontier rows for fifteen minutes, and fills pages from ordered active Mongo pairs without rediscovery. Exact admission always uses normalized keyword units; semantic similarity can rank only that admitted set and EXACT_ONLY never invokes query embedding. A generation-owned ContentSearchRankProfile changes match/semantic/freshness weights independently from the pinned CONTENT_EMBEDDING model, vector dimension, and Qdrant resource. Content-search point UUIDs reserve generation bits while retaining canonical content-key order inside a generation, so Platform can refine a newest-timestamp boundary with Qdrant's stable point-ID order instead of accepting an arbitrary subset of tied rows.

INGEST and ENRICH envelopes carry the tier before model selection; model-backed envelopes carry it only inside ModelExecutionSnapshot. Platform must resolve the logical snapshot before admission, select a physical endpoint per attempt, and never substitute another logical model or tier during retry. The removed physical snapshot fields and prior model_id, error_code, and retry_classification contracts have no compatibility aliases.

Architecture

raw provider HTTP responses
        -> platform-owned adapters
        -> keble-idea schemas and services
        -> platform-owned persistence/projection adapters

MongoDB is canonical in the host application. Qdrant stores canonical concept vectors only, Neo4j stores the concept graph only, Redis stores TTL search frontiers, and RabbitMQ delivers identifiers for Mongo-backed queue envelopes. Ideas are retrieved through scalar idea-to-concept postings rather than idea vectors or multikey story-array scans.

Every active determined posting contributes to a generation-scoped ConceptSearchAvailability row. A posting contributes to the distinctive subset only when both the idea and its matched facet are UNCOMMON. Qdrant and Neo4j mirror both counts on the same canonical concept resource; corpus mode is filtered there and rechecked on Mongo postings. Common-only concepts therefore stay indexed without leaking into seller discovery, while direct key-addressed detail reads remain mode-independent.

Cost-bearing extraction and relationship protocols receive an EnvelopeExecutionIdentity copied from the atomically leased Mongo envelope. The envelope key identifies logical work; attempt_number identifies the physical broker/worker retry. Optional queued_at and leased_at preserve the exact durable queue interval without introducing provider or pricing ownership. Optional bounded forecast_key and catalog-page batch_key preserve typed backfill attribution through normal extraction into Platform operation attempts, without reading untyped envelope domain references. Financial adapters include the identity in provider attempt keys so redelivery cannot overwrite or hide a second possible spend.

ExtractionService is the single batch-identity correction owner. The first typed extraction response may contain missing, unknown, or duplicate content keys; the service issues at most one separately metered corrective request, merges both calls' ProcessingUsage by submitted content key, and rejects a second mismatch with BatchIdentityError. Platform adapters may retain in-run facet validation, but must not add a second identity ModelRetry loop or the application correction and its billing identity become unreachable.

Publication replay has one canonical repository contract: aget_publication_result returns the stored receipt plus the complete idea, concept, and posting-key result for both the initial publish and every receipt replay. ExtractionService never reconstructs a reduced replay result from the receipt alone. Hosts must hydrate this contract from canonical persistence so a paid retry cannot falsely close downstream projection work with empty concept keys.

Side effects if changes:

• initial and replayed extraction return values must remain structurally equal;
• Platform Stage-A work derives its expected projection parts from these keys;
• host reconciliation repairs historical projections separately from replay prevention and must not mutate immutable receipts.

Model identity is a separate immutable value. ModelExecutionSnapshot stores purpose, processing tier, selection rule, logical model/provider identity, catalog fingerprint, and generation without endpoints, credentials, quotas, or prices. EXTRACT, INFER_RELATIONSHIP, PROJECT_CONCEPT, and REINDEX envelopes require their matching purpose; non-model jobs reject the snapshot. INGEST and ENRICH retain an explicit tier until that snapshot exists. Monetary pricing remains owned by the Platform ledger and resolves after one physical endpoint is selected into ProviderAttemptBinding. If no active purpose binding exists, Platform must raise ModelRouteUnavailableError before broker publication. Configuration absence is distinct from a generation-verification or provider-execution failure. Once a route is selected, a temporary transport failure crosses public REST and MCP boundaries only as provider-neutral UpstreamUnavailableError; callers may then explicitly choose a non-provider execution mode instead of receiving a silent fallback. Every new search resolves and freezes one SearchEmbeddingBinding containing the embedding execution snapshot and vector resource. That same binding enters concept retrieval and the cursor session identity, so a host cannot read a model route and Qdrant alias independently.

Development

Python 3.13 and uv are required.

uv sync
uv run pytest -q -m "not live and not slow and not eval and not local_stack and not db_stack and not container"
npx --yes pyright .
uv build

The portable test lane uses only reusable in-memory adapters from keble_idea.testing. Stateful integration adapters belong to the platform and must isolate and clean MongoDB, Redis, Qdrant, Neo4j, RabbitMQ, and artifact state after each test.

Contract rules

• Persist only HIGH and MEDIUM ideas. LOW candidate details are discarded.
• Persist evidence-backed COMMON and UNCOMMON semantics through the same publication path; only UNDETERMINED content omits semantic projections.
• Publish one deterministic content-search profile for every active content classification; do not infer missing profiles from historical receipts.
• Admit catalog results only through exact canonical term units. Semantic similarity is an optional post-admission ranking input, never an admission fallback, and exact-only search incurs no query-embedding attempt.
• Keep catalog rank generations independent from content-embedding generations; rank-weight-only changes rebuild frontiers without rotating vector resources.
• Classify ideas and facets independently, then derive the content summary from the surviving common/uncommon counts rather than trusting a model-owned label.
• Default search to DISTINCTIVE_ONLY; require ALL_DETERMINED explicitly for shopping discovery, and pin the mode into the cursor session filter hash.
• Keep one concept vector and graph node across both modes. A common-only concept remains projected while its determined availability is positive.
• Write exactly one minimal receipt for every content/generation terminal state.
• Hydrate the complete canonical publication result on receipt replay; never return an empty derived-key set for already-published work.
• Keep receipts provider-neutral: physical token/image/duration usage remains in ProcessingUsage, while monetary pricing and per-content allocation belong exclusively to the platform-owned financial ledger. Receipt schema version 4 rejects the removed cost_micros and loose model_id fields.
• Keep batch-identity correction in ExtractionService: one corrective call, one ordinal owned by the host, and one merged receipt usage total.
• Use the eight finite semantic facets; provider/source metadata is not a concept.
• Use standalone_concept_error at model boundaries so invalid facet labels receive a typed correction before publication applies the same strict policy.
• Merge only HIGH-confidence PURPOSE concepts automatically and keep it reversible.
• Vectorize concepts only unless a separately approved retrieval eval changes the design.
• Use immutable generations and active pointers; never mutate active indexes in place.
• Persist named, collection-owned integer-version migration checkpoints so outer workers resume by stable record key and never migrate a corpus at startup.
• Preserve original ASR and OCR independently from English-normalized fields; unknown-language OCR belongs in ocr_text_original, never guessed into ocr_text_en from the transcript language.
• Cutover only verified generations through compare-and-swap pointers and preserve the immediately previous verified generation for exact rollback.
• Report minimum-age or live-session-pin deletion guards as GENERATION_RETENTION_BLOCKED, never as an optimistic revision conflict.
• Stamp relationship coverage with the exact embedding, neighborhood, and relationship generations plus bounded neighbor similarities; a valid null/no-relation result still completes coverage. Persisted current/previous neighbor sets let outer adapters apply entry/exit hysteresis without all-pairs comparison.
• Keep relationship_priority_score and relationship_priority_at as bounded, provider-neutral scheduling hints. The platform derives them idempotently from posting quality/time so Stage-B graph budget reaches high-value/recent concepts first; they never change semantic concept identity or Stage-A search readiness.
• Compact alias postings asynchronously in bounded resumable pages; merge and unmerge correctness must never depend on compaction finishing first. Derived rows retain origin_concept_key, so unmerge removes them without deleting the original posting identities.
• Search is bounded, concept-first, count-free, and cursor-session based.
• Cursor signatures bind expiry, frontier revision, and the complete generation/profile identity.
• Operator lifecycle is canonical and typed (schema version 2): content moves ACTIVE -> ARCHIVED -> ACTIVE and enters PURGING terminally; ideas move ACTIVE -> RETIRED -> ACTIVE and enter PURGING terminally. archived_at / retired_at must match the state, transitions go through assert_lifecycle_transition, and public reads exclude non-ACTIVE rows.
• Operator job commands use the canonical status sets: retry targets TERMINAL_JOB_STATUSES, cancel targets CANCELLABLE_JOB_STATUSES (never LEASED work), and purge work routes through IdeaJobType.PURGE.
• Processing tiers are exactly ECONOMY, BALANCED, and QUALITY. They express requested cost/quality position, not provider identity. A retry keeps the frozen tier; changing it requires explicit reprocessing.

Durable provider recovery

Each queue envelope snapshots the complete active ProviderRecoveryProfileSnapshot, its matching attempt_limit, and an absolute deadline. Automatic retry retains that snapshot. A manual retry also preserves it unless the operator explicitly requests the active profile. Profile records are append-only and activation uses a revision-guarded pointer; changing the pointer affects future admissions only.

The initial profile is fixed at eight attempts, a 72-hour deadline, 30-second fallback doubled to a one-hour cap, a 24-hour maximum provider hint, deterministic 0-10% jitter, five unavailable failures in 60 seconds, a five- minute open circuit, one half-open probe, and blocked release batches of 100. Those are defaults for the first persisted profile, not module constants that silently mutate existing jobs.

retry_at_for_failure anchors delay and jitter to the failure's immutable observation time. Re-evaluating the same envelope attempt therefore returns the same not-before timestamp. A provider hint above the submitted 24-hour policy becomes BLOCKED; attempt/deadline exhaustion or permanent failure becomes DEAD_LETTERED; operator action becomes BLOCKED. Only a retryable failure within both budgets becomes RETRY_SCHEDULED.

JobFailureSnapshot.consumes_attempt_budget is the sole classification for whether a leased retry counts as external work. Queue decision and persistence must use it together: a zero-attempt Platform observation is evaluated against the prior physical-attempt count and its lease increment is refunded atomically. This preserves earlier paid attempts while allowing local projection polling until the immutable deadline.

Restart recovery uses the same envelope identity. QueueLeaseRenewal describes an exact owner/revision/attempt lease extension without implying a domain revision, while ExpiredEnvelopeRecoveryPlan classifies an expired lease as deferred, attempt reuse, fresh-attempt advancement, exhaustion, or a concurrent resolution. A full financial-attempt census supplies the stable attempt keys to abandon; a missing attempt is the only automatic-recovery case that may reuse the queue attempt number. QueueSweepReport is the shared bounded operator result and separates recovered work from live-claim deferrals, exhaustion, and concurrent resolutions.

UpstreamFailureEvidence retains bounded status/code/request identity, opaque quota scope, exact capacity, physical-attempt count, usage, and a completeness bit. Empty usage is known zero only when the producer explicitly marks it complete; incomplete empty usage and every USAGE_ONLY event remain billing unknown. Data Infra evidence also retains a provider-neutral classification origin plus the upstream adapter's mapping revision separately from the Platform mapping revision. This lets operators distinguish provider, contract, and Data Infra runtime classification without making Idea depend on a provider package. Payment requires a reviewed billing-code/account-status source; HTTP 402, negative capacity, quota, and rate limiting are never sufficient by themselves.

Side effects if changes:

• Platform Mongo queue/profile/incident adapters and central indexes implement these exact field paths and compare-and-swap semantics;
• Data Platform job and commerce-operation APIs expose the same failure/policy facts through generated clients and translated frontend labels;
• finance records failed physical usage exactly once while pre-provider gate rejection remains zero spend;
• queue decision and persistence must never duplicate or disagree with consumes_attempt_budget;
• Platform must apply each expired-envelope plan in one queue/financial Mongo transaction and must not clone the envelope or create a parallel status;
• EnvelopeProcessResult must continue to contain exactly one result reference or failure snapshot.

Images are a critical asset

Content payloads treat visuals as first-class evidence: ProductContentPayload.image_artifacts carries model-ready product photos, CommerceProfileContentPayload.image_artifacts carries acquired creator or seller portraits, ShortVideoContentPayload.thumbnail_artifacts carries safe ranked-video thumbnails without transient playback URLs, and VideoContentPayload preserves uploaded video bytes plus key frames. IdeaSearchMedia.from_payload projects the representative image/video (with IdeaSearchSourceSummary for type-specific display facts) onto every search item so consuming platforms can render media-rich cards. Artifact references stay checksum-addressed (ArtifactRef) — delivery URLs are always resolved by the serving platform at read time.

Commerce discovery content

Trending commerce ingestion stays provider-neutral after the platform boundary:

profile = CommerceProfileContentPayload(
    profile_type=CommerceProfileType.CREATOR,
    display_name="Rocket Lab Notes",
    image_artifacts=[acquired_profile_image],
)

video = ShortVideoContentPayload(
    title="A compact model-rocket launch guide",
    thumbnail_artifacts=[acquired_thumbnail],
)

Products keep ProductContentPayload; creators and sellers share the typed profile payload with a stable profile_type; channel-wide ranked videos use a separate short-video payload. Provider names, provider response bags, volatile playback URLs, and raw enum labels do not enter these durable contracts.

Language and standout evidence

ContentBody.original_language remains the backward-compatible ISO code. Optional ContentLanguageProvenance explains whether it was source-declared, text/media detected, or unresolved; a missing provenance on und identifies a legacy row. Provider-neutral ContentStandoutEvidence records why a source surfaced content without claiming profitability. Performance observations are identified by metric kind, unit, observed time, and window: equal duplicates collapse to the standout-linked evidence, while conflicting values are rejected.