Initial public ModelForge release

2026-09-01 21:30:16 +02:00
commit 7082ab955a
490 changed files with 104252 additions and 0 deletions
@@ -0,0 +1,35 @@
+capability: assistant.general
+version: 1
+description: General local text and multimodal assistant workloads.
+input_schema:
+  type: object
+  required: [messages]
+  properties:
+    messages: {type: array, minItems: 1}
+    response_schema: {type: object}
+output_schema:
+  type: object
+  required: [content]
+  properties:
+    content: {type: string}
+    structured_output: {type: object}
+    tool_calls: {type: array}
+modalities: {input: [text, image], output: [text, json, tool_call]}
+languages: [nl, en]
+streaming: true
+structured_output: true
+quality_metrics: [task_success, instruction_following, structured_output_validity]
+upgrade_class: behavioral
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: confidential, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: interactive
+default_residency: keep_warm
+estate:
+  category: ASSISTANTS
+  purpose: Future local general assistant contract; not an M9 executable baseline.
+  stability: planned
+  resource_class: EXCLUSIVE_GPU
+  evaluation_type: llm
+  consumers: [example-ops]
+  payload_limits: {max_bytes: 262144, max_batch_count: 1}
@@ -0,0 +1,24 @@
+capability: assistant.vision
+version: 1
+description: Visual question answering and structured image reasoning.
+input_schema: {type: object, required: [image, prompt], properties: {image: {type: string}, prompt: {type: string}}}
+output_schema: {type: object, required: [content], properties: {content: {type: string}, structured_output: {type: object}}}
+modalities: {input: [image, text], output: [text, json]}
+languages: [nl, en]
+streaming: false
+structured_output: true
+quality_metrics: [task_success, structured_output_validity]
+upgrade_class: behavioral
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: confidential, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: interactive
+default_residency: load_on_demand
+estate:
+  category: ASSISTANTS
+  purpose: Future local image reasoning contract; not an M9 executable baseline.
+  stability: planned
+  resource_class: EXCLUSIVE_GPU
+  evaluation_type: llm
+  consumers: [examplevision, example-ops]
+  payload_limits: {max_bytes: 8388608, max_batch_count: 1, max_width: 4096, max_height: 4096}
@@ -0,0 +1,38 @@
+capability: document.ocr
+version: 1
+description: OCR and structured document extraction.
+input_schema:
+  type: object
+  additionalProperties: false
+  required: [content_base64, media_type]
+  properties:
+    content_base64: {type: string, minLength: 4, maxLength: 11184812}
+    media_type: {enum: [image/png, image/jpeg]}
+    language_hint: {enum: [nl, en, auto]}
+output_schema:
+  type: object
+  additionalProperties: false
+  required: [text, pages]
+  properties:
+    text: {type: string}
+    pages: {type: array, minItems: 1, maxItems: 1}
+    blocks: {type: array}
+    confidence: {type: [number, 'null']}
+modalities: {input: [image, document], output: [text, json]}
+languages: [nl, en]
+structured_output: true
+quality_metrics: [cer, wer, field_accuracy, layout_accuracy]
+upgrade_class: schema_breaking
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: restricted, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: background
+default_residency: load_on_demand
+estate:
+  category: DOCUMENT
+  purpose: Offline text extraction from bounded local document images.
+  stability: experimental
+  resource_class: MEDIUM
+  evaluation_type: ocr
+  consumers: [examplerag, examplevision, example-ops]
+  payload_limits: {max_bytes: 8388608, max_batch_count: 1, max_width: 4096, max_height: 4096, max_pages: 1}
@@ -0,0 +1,40 @@
+capability: rag.embedding
+version: 1
+description: Multilingual text embeddings for semantic retrieval.
+input_schema:
+  type: object
+  additionalProperties: false
+  required: [input]
+  properties:
+    input:
+      oneOf:
+        - {type: string, minLength: 1, maxLength: 8192}
+        - {type: array, items: {type: string, minLength: 1, maxLength: 8192}, minItems: 1, maxItems: 8}
+output_schema:
+  type: object
+  additionalProperties: false
+  required: [vectors, dimension, normalized, embedding_space_id]
+  properties:
+    vectors: {type: array, minItems: 1, maxItems: 8, items: {type: array, minItems: 1024, maxItems: 1024, items: {type: number}}}
+    dimension: {const: 1024}
+    normalized: {const: true}
+    embedding_space_id: {type: string, format: uuid}
+modalities: {input: [text], output: [vector]}
+languages: [nl, en]
+vector: {dimensionality: 1024, normalized: true, cross_deployment_compatible: false}
+slo: {latency_p95_ms: 15000, availability_percent: 99.0, max_concurrency: 1}
+quality_metrics: [recall_at_5, recall_at_10, mrr, ndcg_at_10]
+upgrade_class: requires_reindex
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: confidential, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: production
+default_residency: keep_warm
+estate:
+  category: RAG
+  purpose: Multilingual semantic retrieval for local knowledge projects.
+  stability: stable
+  resource_class: MEDIUM
+  evaluation_type: retrieval
+  consumers: [examplerag, portfolio-assistant]
+  payload_limits: {max_bytes: 65536, max_batch_count: 8}
@@ -0,0 +1,54 @@
+capability: rag.reranking
+version: 1
+description: Second-stage ranking of retrieved text candidates.
+input_schema:
+  type: object
+  additionalProperties: false
+  required: [query, documents, top_n]
+  properties:
+    query: {type: string, minLength: 1, maxLength: 4000}
+    documents:
+      type: array
+      minItems: 1
+      maxItems: 40
+      items:
+        type: object
+        additionalProperties: false
+        required: [id, text]
+        properties:
+          id: {type: string, minLength: 1, maxLength: 255}
+          text: {type: string, minLength: 1, maxLength: 16384}
+    top_n: {type: integer, minimum: 1, maximum: 10}
+output_schema:
+  type: object
+  additionalProperties: false
+  required: [results]
+  properties:
+    results:
+      type: array
+      maxItems: 10
+      items:
+        type: object
+        additionalProperties: false
+        required: [id, score, rank]
+        properties:
+          id: {type: string}
+          score: {type: number}
+          rank: {type: integer, minimum: 1, maximum: 10}
+modalities: {input: [text], output: [score]}
+languages: [nl, en]
+quality_metrics: [mrr, ndcg_at_10, ranking_accuracy]
+upgrade_class: behavioral
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: confidential, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: production
+default_residency: load_on_demand
+estate:
+  category: RAG
+  purpose: Evidence-gated second-stage ranking for retrieval pipelines.
+  stability: experimental
+  resource_class: MEDIUM
+  evaluation_type: retrieval
+  consumers: [examplerag]
+  payload_limits: {max_bytes: 262144, max_batch_count: 40}
@@ -0,0 +1,39 @@
+capability: speech.transcription
+version: 1
+description: Offline bounded speech-to-text transcription.
+input_schema:
+  type: object
+  additionalProperties: false
+  required: [audio_base64, media_type]
+  properties:
+    audio_base64: {type: string, minLength: 4, maxLength: 22369624}
+    media_type: {const: audio/wav}
+    language: {enum: [nl, en, fr, auto]}
+output_schema:
+  type: object
+  additionalProperties: false
+  required: [text, language, duration_seconds]
+  properties:
+    text: {type: string}
+    language: {type: string}
+    duration_seconds: {type: number}
+    segments: {type: array}
+modalities: {input: [audio], output: [text, json]}
+languages: [nl, en, fr]
+streaming: false
+structured_output: true
+quality_metrics: [wer, real_time_factor]
+upgrade_class: behavioral
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: restricted, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: background
+default_residency: load_on_demand
+estate:
+  category: AUDIO
+  purpose: Private local transcription of bounded WAV audio without cloud fallback.
+  stability: experimental
+  resource_class: HEAVY
+  evaluation_type: asr
+  consumers: []
+  payload_limits: {max_bytes: 16777216, max_batch_count: 1, max_duration_seconds: 120}
@@ -0,0 +1,39 @@
+capability: vision.embedding
+version: 1
+description: Image/text embeddings for visual retrieval.
+input_schema:
+  type: object
+  additionalProperties: false
+  required: [items]
+  properties:
+    items:
+      type: array
+      minItems: 1
+      maxItems: 4
+      items:
+        oneOf:
+          - {type: object, additionalProperties: false, required: [image_base64, media_type], properties: {image_base64: {type: string}, media_type: {enum: [image/png, image/jpeg]}}}
+          - {type: object, additionalProperties: false, required: [text], properties: {text: {type: string, minLength: 1, maxLength: 4096}}}
+output_schema:
+  type: object
+  required: [vectors]
+  properties:
+    vectors: {type: array, items: {type: array, items: {type: number}}}
+modalities: {input: [image, text], output: [vector]}
+languages: [nl, en]
+vector: {dimensionality: null, normalized: null, cross_deployment_compatible: false}
+quality_metrics: [recall_at_1, recall_at_5, recall_at_10, mrr]
+upgrade_class: requires_reindex
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: confidential, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: interactive
+default_residency: load_on_demand
+estate:
+  category: VISION
+  purpose: Shared image and text vector space for local visual retrieval.
+  stability: experimental
+  resource_class: HEAVY
+  evaluation_type: visual-retrieval
+  consumers: [examplevision]
+  payload_limits: {max_bytes: 8388608, max_batch_count: 4, max_width: 4096, max_height: 4096}