Entity Enricher MCP Server

Create a reusable benchmark with a mandatory scoring judge. scenario_type='enrichment' needs schema_id and entity_data (the entity to enrich, as enrich_entity takes it — refused when it carries no value; never put it in description); 'sample_generation' needs sample_request; 'schema_generation' needs entity_samples (1..20 samples of one entity type). Enrichment and schema generation need a verified gold reference via set_benchmark_reference before running. Sample generation is rubric-scored and takes no reference. Requires owner and a plan with benchmarks; creating the scenario does not run the models. Returns the scenario and link. Next call run_benchmark when its reference requirements are satisfied. See enricher://docs/model-benchmark.

Delete a benchmark scenario and its stored results. Requires owner and a plan with benchmarks. No LLM call; inspect the scenario before deleting it.

Read one benchmark scenario with per-model quality, cost and speed results. include_reference=true adds the reference, the fixed entity_data and the schema-generation entity_samples. Config changes can make existing results stale. No LLM call. For a ranked subset use get_benchmark_scenario_results. Read after run_benchmark completes.

Filter, rank and limit a scenario's per-model benchmark results. No LLM call. overall blends quality, speed and cost using organization task weights and is null if a component is missing. Status tags are independent: success does not exclude stale or stale_score results. Missing sort metrics come last in either direction. See enricher://docs/model-benchmark for interpretation.

List compact benchmark scenario summaries and total. Scenarios test enrichment, sample generation or schema generation. No LLM call. Read one with get_benchmark_scenario or create one with create_benchmark_scenario. Lifecycle and reference requirements: enricher://docs/model-benchmark.

Start billed asynchronous execution and scoring of a benchmark. Requires owner, a benchmark-enabled plan and a judge; a verified reference is also required except for sample generation. Supply model_keys or providers; omitting both runs all active models with usable provider keys. Repetitions and judging increase cost. Returns job_id and total_models: poll get_job_status, then read get_benchmark_scenario_results. Re-running replaces each selected model's previous result. An organization's benchmark runs and scoring passes execute one at a time: a launch while another is in flight is queued (queue_position, status 'pending'), and a launch on a scenario whose run is still queued folds its models into that run (merged=true, job_id names the queued run). Reference setup and score interpretation: enricher://docs/model-benchmark.

Save the gold reference for an enrichment or schema-generation benchmark. Only set reference_verified=true after checking its values against trusted evidence or obtaining human sign-off; a generated answer alone is not verification. Schema-generation references must be GeneratedJsonSchema objects. Sample-generation scenarios reject references because they are rubric-scored. Requires owner and a plan with benchmarks; no LLM call. A verified reference enables run_benchmark. See enricher://docs/model-benchmark.

Edit a benchmark's test definition or scoring configuration. Requires owner and a plan with benchmarks; no model run. Only supplied fields change, but sample_params and schema_gen_params replace their parameter objects wholesale. The judge may be replaced, not cleared; scenario_type is immutable. Changed test definitions make previous results stale. Re-run affected models to refresh their scores. See enricher://docs/model-benchmark.

Add a property under the root (parent_path=''), an object path or '$defs.X'. Accepts a scalar, nested object or reference to an existing $defs entity or $enums vocabulary. Requires editor; no LLM call. Omitted nullable means required at enrichment. Adding inside $defs affects every usage site. The same validation and working-copy/publication rules as update_schema_property apply. Read the container with get_schema_part first. Property format: enricher://docs/schema-reference.

Generate and auto-save a schema from reviewed samples, returning schema_id, schema content and record links. Supply entity_samples (or samples_csv), sample_record_id, or both; explicit samples override stored JSON while attachment inheritance is preserved. Samples must describe one entity type in one language. Generation combines their fields and annotates relationships; it does not redesign the approved structure. Resolve consequential modeling choices and whether to generate semantic IDs before calling; semantic IDs need an organization embedding model and add cost. Requires editor; synchronous and billed. Review returned suggestions before applying edits with property tools. Sample review and canonicalizations: enricher://docs/schema-from-sample; schema format: enricher://docs/schema-reference.

Soft-delete a saved schema by UUID. Requires editor; no LLM call. Restoration and permanent deletion are available in the web app, not through this tool. Returns the deletion outcome.

Read a saved schema with its properties, annotations and input_contract. Before enrichment, use version='published' for a database-linked schema; default 'working' is for editing. identifying_keys guide entity naming; preserve values belong to the caller and are required; each supplied array item must carry its array_item_keys. Never invent caller-owned values. The response includes version, publish_state and schema_url. A requested published contract that does not exist returns not_found. For a small edit prefer get_schema_part. No LLM call. Contract details: enricher://docs/schema-reference.

Read only the schema fragment needed for an edit. Omit path for the root/type index; use '$defs.X' or '$enums.X' for a definition, an object path for its subtree, or a leaf path for its property card and relations. Dot-separated paths use '[]' for array items. The response identifies shared definition usage, identity participation, database flags and whether entity state or linked databases exist. Editing a $defs property affects every usage site. Reads the working copy; no LLM call. Path examples: enricher://docs/schema-reference.

Edit a saved schema's metadata or replace its full schema_content without an LLM call. Requires editor. Only supplied values change. For one property, prefer update_schema_property, add_schema_property or move_schema_property. Replacements must use GeneratedJsonSchema; unknown keywords are dropped, not rejected, and reported in ignored_keywords / applied_repairs. On database-linked schemas this edits the working copy: neutral edits propagate automatically; structural edits take effect after publish_schema. Returns the updated schema and link. Contract and edit workflow: enricher://docs/schema-reference.

Edit or remove one property by path without replacing the full schema. Requires editor; no LLM call. Only supplied fields change; null clears an entry in flags. Server validation rejects or normalizes invalid combinations and reports applied_repairs. Editing inside $defs affects every usage site. Removing an identity-source member is refused until its identity is recomposed. Renames preserve identity references and record migration intent. Database-linked schemas remain working-copy edits until publish_schema; inspect migration_grade. For additions or relocation use add_schema_property or move_schema_property. Paths, flags and rename rules: enricher://docs/schema-reference.

Register a saved schema for relational synchronization to PostgreSQL, MySQL or SQLite. Requires owner and a sync-enabled plan. Starts a billed classification job when available; returns database ID, classification_job_id or classification_skipped, stamped_keys and registration_notices. The webhook signing secret stays outside the MCP response and can be managed in the web app. The schema is initially unpublished: wait for classification, review keys/options/notices, then publish_schema before any data can sync. pk_strategy locks once the physical model ships. Owned child arrays replace previous membership, so omitted children are deleted on re-enrichment. purge_entity_state transfers custody to replicas; relay custody_warning. A connected host may provision automatically; otherwise get_database_setup_instructions starts browser-confirmed client pairing. Modeling, publication and delivery: enricher://docs/database-sync.

Delete a database registration and its queued deltas, stopping its feed. Requires owner and a sync-enabled plan; no LLM call. External replica tables remain untouched. Entity state and schema database flags remain by default; delete_entity_state and clear_database_model additionally remove data/model settings from schemas left with no registration. Obtain approval for those irreversible teardown options. See enricher://docs/database-sync.

Read the next ordered window of SQL deltas and canonical payloads for a database sync. claim=false is replayable; claim=true leases the window for 120 seconds. Apply a claimed window transactionally, including schema deltas, before ack_database_deltas. limit is also bounded by the registration's page_limit. snapshot_required pauses delivery until the replica reapplies its snapshot. No LLM call. Leasing, cursors, quarantine and snapshot handling: enricher://docs/database-sync.

Return non-secret install, browser-confirmed pairing and run instructions for an ee-database sync client. Requires owner and a sync-enabled plan; no LLM call and no credential is issued or exposed to the MCP client. Run pair_command on the intended replica host: the CLI opens verification_url, the user chooses the database and confirms, and the credential travels directly to the polling CLI. Its DSN stays on that host. Use manual pairing only when managed host provisioning is not already handling the registration. Setup: enricher://docs/database-sync.

List a saved schema's database registrations, linked schemas, options and sync hosts. Returns pending and quarantined delta counts, projection_upgrade_pending and database links. No LLM call. Pending zero alone does not prove healthy delivery: check quarantine and migration blockers. Use list_entity_states for server-side rows and get_record for per-record delivery state. PostgreSQL, MySQL and SQLite delivery is performed by the user's sync client. See enricher://docs/database-sync.

Add an identity concept at zero usage, or add text as an alias using alias_of. Requires editor; resolution may incur embedding/judge cost. Probe first; if the text already resolves to an incumbent, offer that concept instead of blindly retrying. Aliases resolving to another concept are refused. embedding_model selects a new type's space only. Returns concept details and link; manage existing aliases with update_concept_alias. See enricher://docs/semantic-ids.

Delete concepts selected by ids, concept_types or unused_only. Defaults to impact_only=true: review affected records, schemas and replicas before obtaining approval. Execution requires editor; clearing whole types requires owner. Deletion breaks convergence with IDs already stored in replicas; future resolution may mint new IDs. An unused-only deletion still removes that vocabulary. No LLM call. Returns impact counts or deletion results. See enricher://docs/semantic-ids.

Read one concept's aliases, identity source keys, linked records and nearest neighbors within its own type/model slice. Record links are capped; records_truncated signals omitted links. neighbors_limit=0 skips neighbors. Use returned alias IDs with update_concept_alias and similarities to assess a proposed merge. No LLM call. Never compare similarities across embedding spaces or concept types. See enricher://docs/semantic-ids.

Browse organization concepts with aliases, usage counts and type/model facets. view='review' returns pairs escalated by the identity judge, not merely similar pairs. Returns a filtered page; use get_semantic_concept for details and neighbors. Similarities are comparable only within one concept_type/embedding_model slice. No LLM call. Vocabulary review and curation: enricher://docs/semantic-ids.

Analyze sample property ambiguity and relationship identity scoping before schema generation. Requires editor; billed analysis persists a record but does not modify the sample. Returns findings with competing interpretations and suggested_names, plus identity_scoping for sites mixing an entity's own facts with facts about its parent relationship. Names are judged in their parent context, including missing units, periods or ranges. Apply only approved corrections before create_schema_from_sample. Optional: generate_sample already checks its initial sample. How to interpret findings: enricher://docs/schema-from-sample.

Analyze a saved schema's property ambiguity and relationship identity scoping, writing annotations to the schema. Requires editor; billed. Incremental by default; force=true rechecks all sites. Returns findings, suggested descriptions, identity-scoping annotations and pending unification proposals; it does not apply suggested structural changes. Fails with ambiguity_check_disabled if the feature is off. Use update_schema_property for approved descriptions or renames; review structural changes and publish them when linked. Interpretation and modeling guidance: enricher://docs/schema-reference.

Resume a paused job with answers to the questions returned under pause. answers maps question IDs to {option_ids: [...], text: ...}; omitted questions use defaults. Relay consequential unanswered choices to the user unless those defaults were already authorized. Resuming may continue billed model work. Returns the next pause, terminal result or running status after wait_seconds; poll get_job_status if still running. Works with generate_sample clarification in both knowledge and source modes. See enricher://docs/documents.

Answer one LLM request of a job running on your model (every start tool's default unless use_platform_models=true or an explicit platform model): you are its model. Read the request under get_job_status(job_id).llm_requests (or in the start tool's return): run systemPrompt + messages with the listed tools and toolChoice on your own model, then pass the result here as response in MCP CreateMessageResult shape — {role: 'assistant', model: '<your model>', stopReason: 'toolUse' | 'endTurn', content: [{type: 'tool_use', id, name, input}] | {type: 'text', text}} — with optional usage {input_tokens, output_tokens}. toolChoice.mode 'required' means one tool_use block whose input matches the tool's inputSchema. The platform validates the answer like a provider's: a rejected answer returns as a new request with the error in a tool_result (isError) — fix and answer again. Returns the next pending request(s), the terminal result, or the running status after wait_seconds. Not for question pauses (answer_job_question). No platform LLM call. Guide: enricher://docs/llm-delegation.

Fuse two or more records of the same entity into a new arbitration record. Without arbitration_model use voting, median and union rules; with an arbiter, conflicts may incur LLM cost. Returns output, conflicts, fusion metadata and a new record ID. The merged result can feed linked databases and overwrite current entity values. Inspect database warnings and the actual fusion method; an arbiter failure can fall back to rules. Use this after separately recovered model runs, not to merge different entities. See enricher://docs/enrichment-and-fusion. With arbitration_model='auto' YOU arbitrate (unless use_platform_models=true): the call returns the job status at its first pending request (answer_llm_request) or at the terminal state. A rule-based merge (no arbitration_model) makes no model call and blocks until done.

Merge a loser concept into a winner. Defaults to impact_only=true: inspect counts and affected replicas before obtaining approval to execute. impact_only=false requires owner and rewrites aliases, entity identities and referencing payloads, queuing convergence to linked replicas. No LLM call. Similarity alone does not establish identity; inspect get_semantic_concept and the judge review evidence first. Returns impact or merge results. See enricher://docs/semantic-ids.

Read one of your support threads with every turn: your questions (role user), Rupert's answers (assistant) and the support team's follow-ups (admin, with author_name). Opening it marks it read for you, which clears it from threads_with_unread_admin_reply. Continue the thread with ask_support(conversation_id=...). Not found unless the thread is yours (a system admin reads any thread). Guide: enricher://docs/support

Your support threads, newest activity first, with message_count and unread_for_user (a support team follow-up you have not read). Use it to find the conversation_id of an earlier thread to continue or read; a thread you opened in another session is listed too. Guide: enricher://docs/support

Acknowledge every delta through up_to_id after successful application, releasing its lease. Acknowledgement can permanently purge delivered deltas and entity state according to registration options; never acknowledge unapplied or failed work. Returns acknowledged and purged counts. No LLM call. Apply/ack workflow: enricher://docs/database-sync.

Ask Rupert, Entity Enricher's product-support assistant, a question about the application itself: how a feature, a tool or a setting works, what an error_code, a limit, a status or a warning means, why a result or a schema looks the way it does, which plan a feature needs. Use it when the guides (enricher://docs) do not answer, before guessing or telling the user something is impossible. Put the context Rupert needs in the message: the tool called, the error_code or message returned, the schema, job or record concerned (ids are fine, never secrets). Pass conversation_id to continue the thread you opened; omit it for a new topic. Blocks until the answer is complete (up to about two minutes) and returns the full reply. Each question counts against the organization's daily support quota: the result says questions_remaining (null = unlimited) and the call fails with support_question_limit_reached once it is spent, so ask one well-stated question rather than many. A support team member may later post a follow-up in the thread: when the thread you continue holds one you have not read, the result carries the team's follow-ups since your last question as new_admin_messages — relay them to the user, and quote them to Rupert yourself if they matter, since Rupert never reads admin follow-ups; threads_with_unread_admin_reply lists your other threads with one waiting (read them with get_support_conversation). Needs a signed-in user (OAuth or a personal API key): an organization access key gets support_requires_user. attachment_ids are ordinary uploads (upload_attachment): images up to 2 MB and text documents only. Guide: enricher://docs/support

Assign or clear the host provisioning a database sync. Requires owner and a sync-enabled plan. host accepts a connected host ID/name; null unassigns it. Assignment may create the physical database and start synchronization automatically. Moving hosts revokes the old host's minted credential; an existing manual pairing is not evicted. Candidates come from create_database_sync or list_database_syncs. No LLM call. Managed and manual setup: enricher://docs/database-sync.

Permanently delete an attachment in your organization, including its stored file. No LLM call. Existing records remain, but future calls or schema regeneration cannot reuse the deleted source. Delete only when that source is no longer needed; this is not a required post-enrichment step. Returns the deleted ID and filename.

Request cancellation of a pending, running or paused LLM job. In-flight model calls may finish and persist records; remaining work is skipped. Cancellation does not roll back records or database writes. Read get_job_status and list_records(job_id=...) afterward to inspect the outcome. No new LLM call is started by this tool.

Start a billed analysis proposing database keys, SQL types, indexes and relationship ownership on a linked schema. Requires editor. Registration already starts the initial pass; use this after relevant edits. Incremental scope covers new properties or changed JSON types/multilingual flags; an empty scope does not rerun unchanged fields. Returns job_id: poll get_job_status and inspect get_schema's working copy. Correct proposals with property tools, then publish_schema to ship changes. Entity-level indexes and ownership choices: enricher://docs/database-sync.

Remove or promote a concept alias using alias IDs from get_semantic_concept. Requires editor; no LLM call. action='remove' stops that surface form resolving to this concept; removing the last alias is refused. action='set_canonical' changes its displayed form. To add an alias use add_semantic_concept(alias_of=...). Returns the outcome. See enricher://docs/semantic-ids.

Enrich one entity against exactly one of schema_id or target_schema, returning structured output, record_id, costs and any database outcome. Read get_schema's input_contract first (published version when linked); never invent preserve values. Models may be omitted for auto selection. Generation is billed. Two or more models fuse only when all succeed: check failed_models before reporting success. A failed leg prevents automatic fusion and database admission. classification_warning returns success=false, error_code and classification; bypass only after user confirmation. Database sync defaults on: report database.status and database_warning, including partial writes or overwrites; admission is not proof of replica delivery. Recovery and fusion: enricher://docs/enrichment-and-fusion. By default the job runs on YOUR model (use_platform_models=false, no explicit model): the call returns the job status as soon as the first model call waits for you (llm_requests), or at the terminal state; a classification warning then pauses the job for answer_job_question instead of returning classification_warning, and enable_web_search must be decided (omitted returns web_search_undecided). On platform models the call blocks until completion.

Browse a schema's current merged entity rows, not per-run records. Requires editor; no LLM call. Returns identities, revision, last_record_id and optionally payload, with limit/offset pagination. Rejected entities have no row; purge_entity_state may also remove delivered rows. Server-side state does not prove the external replica has applied its deltas. Use get_record and list_database_syncs for delivery and rejection diagnostics. See enricher://docs/database-sync.

List observed values outside each open enum's current vocabulary, with counts from recent enrichment records. No LLM call. Use the report to propose admitted members or rejected_values, or close a vocabulary only when it is exhaustive. This read does not edit the enum. An open enum allows other values; a closed one constrains output to its members. Named enums can be read with get_schema_part and changed through update_schema. See enricher://docs/schema-reference.

Generate editable sample JSON from a free-text request for schema authoring. Without attachments, use model knowledge and optional web search; with attachments, extract from the sources only (search does not relax that rule). Each sample is one instance in one language; set sample_count separately from request. Attachments force one sample. Requires editor; generation is billed. Returns a job_id and may already be paused or complete: relay pause questions through answer_job_question, otherwise poll get_job_status. Review returned samples and warnings before create_schema_from_sample; do not silently change facts or structure. For relationship modeling, multiple documents or hybrid extraction plus research, read enricher://docs/schema-from-sample and enricher://docs/documents.

Resolve 1..1000 texts against one concept type. mint=false returns exact/matched/would_mint outcomes without minting; mint=true creates misses and requires owner (reporting requires editor). Resolution can call embeddings and the identity judge even in report mode. Review would_mint rows before authorizing creation. A new type may be initialized with embedding_model; existing types cannot switch spaces through import. Returns per-text outcomes. See enricher://docs/semantic-ids.

Read a job's status, progress and compact terminal summary with persisted record IDs. Status is pending, running, paused, completed, failed or cancelled. Relay pause questions with answer_job_question; answer a delegated job's llm_requests with answer_llm_request. include_result=true returns full terminal details; batch summaries count entities and database outcomes. A failed job is not usable output; inspect individual records for partial recovery. Unknown IDs may be invalid, expired or lost after restart: look for persisted outputs with list_records(job_id=...), without assuming success. events_after=<seq> adds the job's event log past that cursor (per-model completions, scoring progress, pauses) — the poll equivalent of the SSE stream; pass the returned last_seq next time. On a delegated job, llm_requests lists the pending model calls: the first complete, the rest as envelopes; llm_request_id=<id> returns that one complete. No LLM call. Polling, failure and recovery guidance: enricher://docs/enrichment-and-fusion.

Inspect or migrate the organization's concept embedding space. action='status' is read-only; preview, start and cancel require owner. preview estimates cost and reports potential collisions; review them before authorizing start. target_model is required for preview/start; source_model and concept_types scope the move. Starting launches billed background re-embedding while enrichment continues, then switches the selected slices after coverage is complete. Cancel marks the transition cancelled; it does not guarantee interruption or rollback of in-flight work. Returns migration status or preview. See enricher://docs/semantic-ids.

List available model keys, nominal capabilities, languages, strategies, auto-selected defaults and organization profile_limits. Use when choosing explicit models or checking plan limits; ordinary calls may omit models or use auto without fetching this large catalogue. Auto also accounts for attachment capabilities. is_available means a usable provider key exists, not that provider quota or every combination of tools and media will work. Missing capability flags mean unsupported on this discovery surface. default_models_web_search is a search-only preview, not attachment-specific. No LLM call. Model selection and costs: enricher://docs/enrichment-and-fusion.

Move one property into the root, an object path or '$defs.X', preserving its flags and expertise. Requires editor; no LLM call. Moving into a definition changes every usage site. An identity composition naming the property is re-spelled automatically when the destination stays within the same 1-1 closure (identity_rebound reports it); the move is refused when it would take the key out of that closure or leave a semantic ID with no identity material, and on collisions or recursive containment. Read source and destination with get_schema_part first. Structural changes on database-linked schemas require publish_schema. Paths and migration guidance: enricher://docs/schema-reference.

Materialize an entity region from get_schema's x-entityMap. Ordinary flat members (e.g. product_id, product_name on an order line) move into a new object named after the region, regions hanging from it move along (nest them in turn inside the new object), pairing facts stay on the host, and the moved names shed the region's tokens (product_name → name) unless strip_prefix=false. A pairing region nests the same way: its pair facts gather into a pairing object with the entity hanging from it (series_volume → series_engagement.volume beside book_series). Compact scalar occurrences (e.g. manufacturer_name or each therapeutic_classes[] item) are all converted to references to one shared $defs entity; host_path and strip_prefix do not apply to that form. If the schema already uses semantic IDs, the created object gets one when generation's gates allow (keyed, not only by exact values, not a pairing); no other object changes. host_path is '' for the root, an object path, 'path[]' for an array's items, or '$defs.X'. Defaults to dry_run=true: inspect the returned schema_content and notes, then persist with dry_run=false only after approval of the change. Requires editor; no LLM call. Database-linked structural edits still require publish_schema. Modeling consequences: enricher://docs/schema-reference.

Preview identity resolution without adding a concept or increasing its usage. Requires editor; uncached resolution may call embeddings and the identity judge. Returns exact_hit, match or no_match, the matched concept and neighbors. Probe before adding; a matched incumbent may already represent the intended entity. embedding_model can select the space for a new concept type, not change an existing type's space. Inspect judge evidence as well as similarity. See enricher://docs/semantic-ids.

Publish a database-linked schema's working copy as the contract used by enrichment and replicas. A newly linked schema sends nothing until first publication; unlinked drafts cannot be published. Requires editor; no LLM call. Call validate_only=true to inspect the diff, blockers, warnings and per-database migration_sql. Transform migrations require user-approved confirm_transforms=true; cross-schema conflicts remain blockers. key_language may be required if classification revealed a multilingual key. Returns publication state; queued migrations apply asynchronously on replicas. See enricher://docs/database-sync for the review and delivery workflow.

Read one persisted record's structured_output, entity_input_data, validation errors, expertise verdicts and metrics. failed_expertises and partial identify incomplete work even when output exists; use retry_expertises only on a record with failed domains. Fusion metadata identifies its source models and arbitration method. database_sync, when present, reports admission and per-replica delivery state. Returns record_url and the related schema link. No LLM call. Interpretation and recovery: enricher://docs/enrichment-and-fusion.

List compact, paginated records in your organization, most recent first. Filter by job_id to retrieve persisted outputs of an asynchronous workflow; type, success, model and search further narrow the result. Both per-model enrichment and arbitration records may exist for one entity. Use get_record for full output and failures; use list_entity_states for current merged entity state instead of run history. No LLM call. Fetch every page when a job produces more than one page of records.

Resolve one pending entity-type unification proposal from get_schema. action='accept' maps the proposed site onto the winning $def; field_map overrides proposed correspondences. Unmapped fields are retained as nullable; other usages of the winning definition are affected. action='dismiss' keeps the sites separate. Defaults to dry_run=true: inspect the returned schema_content and notes, then persist with dry_run=false only after approval of the change. Requires editor; no LLM call. Database-linked structural edits still require publish_schema. Modeling consequences: enricher://docs/schema-reference.

Retry only an existing record's failed expertise domains, then update its output and attempt the run's fusion/synchronization. Billed for retried work. Requires failed_expertises on that record (get_record); a surviving successful sibling is not retryable and returns no_failed_expertises. Supply its entity_input_data and saved_schema_id; model optionally substitutes the failed model. Returns job_id: poll get_job_status, then re-read get_record. When the failed leg left no record, use a one-model enrich_entity with database_sync=false followed by merge_records instead. Recovery decisions: enricher://docs/enrichment-and-fusion.

Undo automatic edits a scoring pass made to a scenario's reference. Every scoring pass folds what the scored models showed the reference should be (a candidate the judge found better, a rule the samples prove, a value the reference lacked) and writes it into the reference — an edit a pass already made is only replaced by stronger evidence (samples, a wrong verdict, more agreeing models), never by one more model's better verdict; the log is reference_meta.auto_applied on get_benchmark_scenario, each entry with its inverse patch. Pass the revision ids to undo: the inverse is applied, the entry is marked reverted and its (path, attribute) is pinned so no later pass re-applies it (a manual set_benchmark_reference lifts every pin). Scores never go stale from this. Requires owner and a plan with benchmarks. No LLM call.

Save a directly authored schema and return its ID and link. Requires editor; no LLM call or generation charge. schema_content is a JSON Schema 2020-12 document in Entity Enricher's supported dialect: title, type='object', properties, optional $defs and x-* extension sections. The server validates it and makes colliding names unique. Entity definitions, enum vocabularies and localized fields have different projection rules. Unknown keywords are dropped, not rejected: read ignored_keywords (path, keyword, hint) and applied_repairs in the result — a property flag such as semantic_id placed on a $defs entity object lands there, with the level it is read at. Use create_schema_from_sample to derive a schema from data, or update_schema for an existing schema. A minimal valid example and supported annotations are in enricher://docs/schema-reference.

List saved schemas in your organization, pinned first. Returns compact summaries with IDs, names and links; no LLM call. Use get_schema to inspect a chosen schema and its input contract, or get_schema_part for a targeted edit. For a database-linked schema, enrich against its published version; the working copy may contain unpublished changes.

Start billed asynchronous enrichment of an entity list against exactly one of schema_id or target_schema. Returns job_id and total. No fixed entity-count cap; live prompt quotas and credits can stop remaining work. Each entity follows the enrichment/fusion pipeline; every model must succeed for its automatic fusion and database admission. A confident classification mismatch skips that entity without enrichment, never pauses the batch. Poll get_job_status, then list_records(job_id=...). Attachments apply to every entity. Unlike enrich_entity, this tool does not expose database_sync=false. By default the job runs on YOUR model (use_platform_models=false, no explicit model): its entities park their calls as llm_requests, several at once, and enable_web_search must be decided (omitted returns web_search_undecided). Input contracts, partial results and recovery: enricher://docs/batch-enrichment.

Read organization-wide record totals, success rate, tokens and cost summary. No LLM call. This tool has no job filter; use list_records(job_id=...) for a particular run and benchmark tools for comparative quality/cost/speed scores.

Validate and inject stored or supplied enrichment output into the entity layer and linked syncs. May incur semantic-resolution cost. record_id alone reuses its output; adding structured_output creates a new derived record. Without record_id, supply structured_output and saved_schema_id. Each item uses the current published contract and admission gate. Only enrichment/arbitration records qualify; base records whose fusion is in the same request are skipped. Inspect each outcome for rejected or partial writes. Use after fixing a rejected output or an enrich_entity run with database_sync=false. See enricher://docs/enrichment-and-fusion.

Upload base64 file bytes as reusable source material; returns id and requires_capability. Pass the ID in attachment_ids to sample/schema generation, enrichment or benchmarks. Supported format handling depends on server MIME policy: extracted text or model-readable binary. Prefer auto model selection for attachment capabilities. Uploading does not itself run an LLM. A generated sample with attachments is source-only; see enricher://docs/documents for formats, multiple-file behavior and research workflows. Retain attachments needed for later runs or regeneration.