Files

Eric Rakestraw a263a0840c Minimize D&D NPC extraction contracts

2026-07-22 18:51:26 +00:00

28 KiB

Raw Blame History

Configuration

This is the canonical reference for implemented Notarius configuration.

Notarius reads YAML config files with version: 3. File configuration is applied over built-in defaults, then environment overrides are applied. Explicit CLI overrides are applied last where the command supports them.

Discovery

Commands that load configuration use this order:

an explicit path supplied through the CLI, when provided;
NOTARIUS_CONFIG, when set to a non-empty path;
/usr/local/etc/notarius/config.yml.

If none is available, the command fails with a config file not found error. The explicit-path option is defined in the CLI reference.

Maintained Examples

All are complete version 3 files. The fragments below illustrate individual fields and are not alternate complete configurations.

Top-Level Fields

version: required. The only supported value is 3.
scriptorium: optional Scriptorium profile source settings.
pipelines: optional map of pipeline IDs to pipeline definitions.
concurrency: optional global concurrency settings.
output: optional durable output placement.
cache: optional chunk-plan and checkpoint cache placement.
debug: optional debug-bundle placement. It does not enable debug capture.

Unknown YAML fields are rejected. The removed top-level llm_profiles field is rejected; execution profiles now come from Scriptorium.

Defaults

Built-in defaults:

concurrency.total_llm: 1
concurrency.stage_workers.extract: effective concurrency.total_llm
output.directory: ./notarius-output
cache.chunk_plans.mode: auto
cache.chunk_plans.directory: unset, selecting <os.UserCacheDir>/notarius/chunk-plans
cache.checkpoints.enabled: false
cache.checkpoints.directory: unset, selecting <os.UserCacheDir>/notarius/checkpoints
debug.directory: ./notarius-debug

No pipelines are built in. A run requires a configured pipeline.

If scriptorium is omitted, Notarius uses Scriptorium's built-in profile catalog. Prompt definitions may also name default profile IDs. The current D&D scene and spell prompts default to the built-in mistral-small-3 profile when a module binding does not set llm_profile. That built-in profile reads its credential from OPENROUTER_API_KEY.

Scriptorium Profiles

scriptorium fields:

profile_dir: optional directory containing Scriptorium profile YAML files.
profile_file: optional Scriptorium profile YAML file.

profile_dir and profile_file are mutually exclusive. Custom profiles overlay Scriptorium built-in profiles by profile ID.

Scriptorium profile files use Scriptorium's profile schema. A minimal profile looks like:

id: local-fast
endpoint: http://127.0.0.1:8080/v1
model: your-model
api_key_env: SCRIPTORIUM_API_KEY
timeout_seconds: 180

Notarius does not accept raw API keys in Notarius config. For file-backed Scriptorium profiles, store the environment variable name in api_key_env and set that variable in the run environment. Scriptorium rejects raw api_key fields in profile YAML.

Environment Overrides

These environment variables are applied after the config file:

NOTARIUS_CONFIG: config discovery path.
NOTARIUS_TOTAL_LLM_CONCURRENCY: integer global LLM concurrency.
NOTARIUS_STAGE_WORKERS_EXTRACT: integer extract worker limit.
NOTARIUS_OUTPUT_DIR: durable output root.
NOTARIUS_CACHE_CHUNK_PLANS_MODE: chunk-plan cache mode.
NOTARIUS_CACHE_CHUNK_PLANS_DIR: chunk-plan cache root.
NOTARIUS_CACHE_CHECKPOINTS_DIR: checkpoint cache root.
NOTARIUS_DEBUG_DIR: debug-bundle root.

Integer environment values must parse as base-10 integers. Directory overrides must be non-empty after trimming. Cache-directory fields in a file may be empty, which deliberately selects the corresponding per-user default.

The removed NOTARIUS_LLM_DEFAULT_* variables are not read. Configure provider endpoint, model, and credential environment variable names through Scriptorium profiles.

Concurrency

concurrency fields:

total_llm: positive integer ceiling on concurrent provider calls.
stage_workers: optional map of framework worker limits. The only supported key is extract.

stage_workers.extract defaults to the effective total_llm value after file and environment precedence. It must be between 1 and total_llm, inclusive. Unknown or empty stage-worker keys are rejected. The environment override NOTARIUS_STAGE_WORKERS_EXTRACT takes precedence over the file value, as does NOTARIUS_TOTAL_LLM_CONCURRENCY for the global ceiling.

The worker value is present in effective and redacted configuration. It bounds the fixed run-wide extract pool and its bounded dispatch queue. Extract jobs are submitted by source chunk and then resolved lane; total_llm independently bounds actual provider calls made by extracts, retries, and validators.

Pipelines

A pipeline selects implementations for the fixed workflow defined by Architecture.

Pipeline fields:

input: required module binding.
chunk: optional module binding. Default module is generic.
artifacts: the artifact lane map for a single-step pipeline. It is treated as an implicit step with the stable ID default.
steps: an ordered, non-empty list of step definitions. A pipeline may use steps or artifacts, but not both. Step IDs must be unique after trimming.
output: optional module binding. Default module is json.
references: optional map of reference slot names to reference paths. These bindings are defaults for eligible pipeline targets that declare the matching slot.

Each explicit step contains an id, an optional references map, and a non-empty artifacts map. Steps share the pipeline input, chunk plan, worker budget, output encoder, manifest, and failure boundary. Lanes within a step retain the fixed extract, validate, merge, validate, normalize, and validate workflow; the next step starts only after the current step is terminal.

Generated references use the structured artifact source form to identify one accepted normalized lane from an earlier step:

pipelines:
  dnd-npc-grounded:
    input: seriatim
    steps:
      - id: identify-npcs
        artifacts:
          npcs:
            extract: dnd/npcs
            normalize: dnd/npcs
      - id: grounded-events
        references:
          npcs:
            artifact:
              step: identify-npcs
              lane: npcs
        artifacts:
          spells:
            extract: dnd/spells
            normalize: dnd/spells
          combat:
            extract: dnd/combat-turns
            normalize: dnd/combat-turns

The generated binding is explicit and typed; it is not inferred from module, lane, or slot names. It may be declared at step scope, applying to every selected target in that step that declares the slot, or at one target's references map. A producer may fan out to compatible target slots, but a slot accepts only one producer. A producer must be in an earlier step, and a configured generated dependency is required even when the consumer slot is otherwise optional. Aggregating several producer artifacts is unsupported.

The producer codec supplies the artifact kind, complete schema identity, media type, canonical content digest, and size used for compatibility and checkpoint dependency checks. Only one accepted normalized artifact crosses the boundary; raw extraction results, rejected output, intermediate values, and validator diagnostics do not. Generated content is supplied in memory and is never represented by a filesystem path.

Artifact lane fields:

extract: required module binding.
merge: optional module binding. Default module is appendorder.
normalize: optional module binding. Default module is noop.
validators: deprecated lane-level validator list. Non-empty lists are rejected; use extract.validators, merge.validators, or normalize.validators.
references: optional compatibility alias for extractor reference bindings. Lane bindings override pipeline-level bindings for the same slot.

Commands that resolve a pipeline fail for unknown or incompatible module keys. See CLI Reference for command syntax.

Reference bindings are validated against reference slots declared by eligible chunk, extract, merge, and normalize targets during pipeline resolution. Required slots must be bound after config defaults, target-local references, lane-level compatibility bindings, step-local references, and command-line reference overrides are applied. Config-relative paths are resolved relative to the config file; command-line reference paths are resolved relative to the current working directory. Bound files must be UTF-8 text. Reference media types are inferred from file extensions and checked when a module restricts accepted types; unknown extensions use application/octet-stream. See CLI Reference for command-line selectors and Operations for recorded provenance and sensitive-data handling.

Pipeline-level references are defaults. They are valid when at least one eligible target in the full configured pipeline declares the slot, including chunk, extractor, merger, and normalizer targets. During a run, they apply only to the selected targets that declare the slot. For external bindings, a step-local binding overrides a pipeline-level default and a target-local binding retains the existing most-specific precedence. Generated and external bindings may not resolve to the same target slot, and a step-scoped generated binding may not be duplicated by a target-local generated binding.

pipelines:
  dnd-session:
    input: seriatim
    references:
      players: ./campaign/players.txt
      party: ./campaign/party-roster.txt
      glossary: ./campaign/glossary.txt
    artifacts:
      spells:
        extract: dnd/spells

Extractor binding references are the canonical lane-local location. The legacy lane-level references field remains supported as an alias; when both bind the same slot, extract.references wins:

pipelines:
  dnd-session:
    input: seriatim
    references:
      glossary: ./campaign/glossary.txt
    artifacts:
      spells:
        references:
          roster: ./campaign/legacy-roster.txt
        extract:
          module: dnd/spells
          references:
            party: ./campaign/session-party.txt

chunk.references, merge.references, and normalize.references are accepted in object-form bindings. They override pipeline-level defaults for slots declared by that target module. Extractor-local references apply only to the extractor, merger-local references apply only to the merger, and normalizer-local references apply only to the normalizer.

Target-local reference fields use the same map shape at:

pipelines.<id>.chunk.references
pipelines.<id>.artifacts.<lane>.extract.references
pipelines.<id>.artifacts.<lane>.merge.references
pipelines.<id>.artifacts.<lane>.normalize.references

Each binding is valid only when that target module declares the slot.

Reference source forms are:

a scalar string, which is an external file path; or
an object with only artifact, containing trimmed step and lane IDs for an earlier producer lane.

Pipeline-level references accept only external paths. Generated references are valid at step scope or on a target-local references map. Their producer's registered codec is authoritative for schema, media type, and canonical content identity; an external file is not treated as generated merely because its bytes decode as the same artifact.

Module Bindings

Every module binding may use shorthand:

input: seriatim

or object form:

chunk:
  module: dnd/scenes
  llm_profile: local-fast

Binding fields:

module: module key.
llm_profile: optional Scriptorium profile ID. Empty or omitted lets the Scriptorium prompt default select the profile.
retries: non-negative retry count for extra runtime attempts after the first attempt. Default: 0. Supported on chunk, extract, merge, and normalize bindings.
options: optional module-specific settings.
references: optional reference bindings. Supported only for chunk, extract, merge, and normalize bindings. input and output bindings reject this field during validation.
validators: optional stage-local validator chain override. Supported only for chunk, extract, merge, and normalize bindings. Omit the field to use the production default chain; set validators: [] to force an empty chain; set a non-empty list to use exactly those validators in configured order.

During resolution, each selected module's registered option validator runs. Production input, chunk, and output bindings reject unknown or invalid options with the affected binding context.

Validator bindings use the same shorthand or object module-binding form, but only these fields are supported:

module: validator key.
llm_profile: optional Scriptorium profile ID for LLM-backed validators.
options: optional validator-specific settings.

Validator bindings reject references, retries, and nested validators. During resolution, deterministic validators reject explicit llm_profile values.

Configured LLM-backed validators with explicit llm_profile values are validated against the configured Scriptorium profile source. Deterministic production validators do not call the LLM and must not set llm_profile.

Implemented Production Modules

Slot	Key	Notes
input	`seriatim`	Reads Seriatim transcript JSON.
chunk	`generic`	Splits source units into ordered chunks.
chunk	`dnd/scenes`	Uses an LLM to split transcript source units into D&D scenes.
extract	`dnd/spells`	Extracts typed D&D spell-list artifacts.
extract	`dnd/npcs`	Extracts typed D&D NPC-list artifacts.
extract	`dnd/combat-turns`	Extracts typed D&D combat-turn-list artifacts.
merge	`appendorder`	Combines typed artifacts in chunk order.
normalize	`noop`	Passes merged typed artifacts through unchanged.
normalize	`dnd/spells`	Deterministically canonicalizes and de-duplicates typed D&D spell-list artifacts.
normalize	`dnd/npcs`	Deterministically consolidates typed D&D NPC-list artifacts by canonical name and unions exact evidence.
normalize	`dnd/combat-turns`	Deterministically canonicalizes, orders, and de-duplicates typed D&D combat-turn artifacts.
output	`json`	Produces JSON output files for normalized `application/json` lanes.

Implemented Production Validators

Key	Execution	Notes
`generic/always_accept`	deterministic	Accepts returned module output.
`generic/always_reject`	deterministic	Rejects returned module output with reason `always_reject`.
`generic/valid_json`	deterministic	Rejects payloads that are not syntactically valid JSON.
`generic/valid_json_schema`	deterministic	Rejects invalid JSON or JSON that does not conform to the module response schema.
`extract/dnd/spells/shape`	deterministic	Rejects malformed D&D spell-list artifacts.
`extract/dnd/spells/catalog`	deterministic	Rejects spell-list artifacts containing names outside the effective SRD and overlay catalog.
`extract/dnd/spells/source_refs`	deterministic	Rejects missing or invalid D&D spell source references.
`extract/dnd/spells/source_relatedness`	deterministic	Emits warnings when a spell name is not found near its cited source text.
`extract/dnd/npcs/shape`	deterministic	Rejects malformed D&D NPC-list artifacts.
`extract/dnd/npcs/source_refs`	deterministic	Rejects missing or invalid D&D NPC source references.
`extract/dnd/npcs/source_relatedness`	deterministic	Emits warnings when an NPC name is not found near its cited source text.
`normalize/dnd/npcs/identity`	deterministic	Rejects invalid canonical IDs and duplicate canonical-name or ID ownership.
`extract/dnd/combat-turns/shape`	deterministic	Rejects malformed D&D combat-turn artifacts.
`extract/dnd/combat-turns/source_refs`	deterministic	Rejects missing or invalid D&D combat-turn source references.
`extract/dnd/combat-turns/source_relatedness`	deterministic	Emits warnings when an actor or declared action is not found near cited source text.
`normalize/dnd/combat-turns/invariants`	deterministic	Rejects normalized combat-turn identity, target, evidence-order, and chronology violations.

The production default chain for dnd/spells is used for both its extract and normalize stages:

validators:
  - generic/valid_json
  - generic/valid_json_schema
  - extract/dnd/spells/shape
  - extract/dnd/spells/catalog
  - extract/dnd/spells/source_refs
  - extract/dnd/spells/source_relatedness

The production default chain for dnd/npcs uses the extraction chain for the extract stage and the identity chain for normalize-stage output:

extract:
  validators:
    - generic/valid_json
    - generic/valid_json_schema
    - extract/dnd/npcs/shape
    - extract/dnd/npcs/source_refs
    - extract/dnd/npcs/source_relatedness
normalize:
  validators:
    - generic/valid_json
    - generic/valid_json_schema
    - normalize/dnd/npcs/identity
    - extract/dnd/npcs/source_refs
    - extract/dnd/npcs/source_relatedness

The production default chains for dnd/combat-turns are:

extract:
  validators:
    - generic/valid_json
    - generic/valid_json_schema
    - extract/dnd/combat-turns/shape
    - extract/dnd/combat-turns/source_refs
    - extract/dnd/combat-turns/source_relatedness
normalize:
  validators:
    - generic/valid_json
    - generic/valid_json_schema
    - extract/dnd/combat-turns/shape
    - normalize/dnd/combat-turns/invariants
    - extract/dnd/combat-turns/source_refs
    - extract/dnd/combat-turns/source_relatedness

Empty chains approve output by default.

The generic chunker accepts:

max_units: positive integer, default 50;
overlap_units: non-negative integer, default 0, and must be less than max_units.

The dnd/scenes chunker requires transcript source capabilities, calls the configured structured LLM runtime, and does not accept module options. It declares optional players, party, and glossary references for scene disambiguation, and accepts roster as a deprecated compatibility alias for party.

The dnd/spells extractor declares optional campaign reference slots:

players
party
glossary
roster as a deprecated compatibility alias for party

These campaign slots accept UTF-8 plain text, Markdown, YAML, or JSON reference files. The extractor also declares an optional spell_catalog slot that accepts one UTF-8 application/json overlay bundle no larger than 1 MiB. The slot does not allow multiple files. Its format is defined in the spell-catalog overlay contract. The extractor uses campaign references only as supporting disambiguation material; spell casts still must be present in the source transcript.

It also declares an optional npcs slot for a normalized NPC artifact. The slot accepts exactly one application/json artifact no larger than 1 MiB. An external file is decoded and identity-validated during preparation. A generated binding is validated at the step handoff and is provided to the operation through the same reference contract. In both cases, the model receives a names-only JSON projection for caster-name grounding. Registry source references may belong to the NPC-producing session and are provenance only; they are not spell evidence. Generated reference identity and bounded producer provenance are recorded by the framework; NPC names, content, and paths are not copied into manifests. Consumer-local checkpoint identity uses the names-only projection digest. When absent, the prompt receives the exact empty value {"npcs":[]} with its projection digest and no registry provenance.

The dnd/spells normalizer declares the same optional spell_catalog slot. When an overlay is used, bind it independently under artifacts.<lane>.normalize.references.spell_catalog; normalize-stage references are local to that stage and are not inherited from extraction. The normalizer uses the embedded SRD catalog when no normalize-stage overlay is bound.

The dnd/npcs extractor declares the same optional campaign slots as the spell extractor, but it does not declare the npcs registry slot. Its normalizer accepts no references. The maintained NPC-grounded example binds its accepted normalized output to the later spell and combat targets through an explicit ordered step.

The dnd/combat-turns extractor declares the optional campaign slots and the structured npcs slot. Campaign references guide only the LLM extraction stage. The deterministic normalizer declares only npcs, whose operation-time registry supports the same actor and target canonicalization. Each npcs slot accepts exactly one UTF-8 application/json artifact no larger than 1 MiB. The registry's source ranges remain provenance for the reference and never become combat evidence. An ordered step binding fans the same generated NPC artifact out to extraction and normalization:

references:
  npcs:
    artifact:
      step: identify-npcs
      lane: npcs

When bound, the combat extractor and normalizer receive the generated registry at operation time. Framework provenance and checkpoint dependencies contain its kind, schema identity, media type, canonical digest, size, and bounded producer identity; names, content, and paths are not recorded there. When absent, the combat prompt receives the exact empty registry value {"npcs":[]} with its projection digest and no registry provenance.

State Surfaces

The output, cache, and debug top-level fields select independent physical roots. Their layout, permissions, lifecycle, and sensitive-data handling are defined in Operations.

output:
  directory: ./notarius-output
cache:
  chunk_plans:
    directory: ""
    mode: auto
  checkpoints:
    enabled: false
    directory: ""
debug:
  directory: ./notarius-debug

output.directory is the durable output root. Its precedence is --output-dir, NOTARIUS_OUTPUT_DIR, the file value, then the default.

cache.chunk_plans.mode accepts auto, bypass, or refresh. Its precedence is --chunk_cache, NOTARIUS_CACHE_CHUNK_PLANS_MODE, the file value, then auto. auto reuses a valid source-addressed plan and regenerates missing or invalid records; bypass performs no plan-cache I/O; refresh regenerates and publishes a plan after chunk validation.

cache.chunk_plans.directory and cache.checkpoints.directory each name an exact cache-family root. Their precedence is the corresponding environment variable, the file value, then the family-specific per-user default. There is no CLI cache-root override. The defaults are <os.UserCacheDir>/notarius/chunk-plans and <os.UserCacheDir>/notarius/checkpoints; on Unix, os.UserCacheDir ordinarily uses an absolute $XDG_CACHE_HOME or falls back to $HOME/.cache. A relative XDG_CACHE_HOME is an error.

cache.checkpoints.enabled defaults to false. When true, every run records checkpoint transitions and reusable approved results. When false, Notarius does not resolve or create the checkpoint root, and --resume is rejected. The --resume flag authorizes loading compatible checkpoints; it does not control recording.

debug.directory chooses a root but never enables debug capture. Its precedence is --debug-dir, NOTARIUS_DEBUG_DIR, the file value, then the default. Only --debug requests a bundle; --debug-dir is valid only with --debug.

Every supplied file, environment, and CLI value is validated even when a higher-precedence value wins.

Version 2 To Version 3 Migration

Version 2 files are rejected. Move each setting to the surface it controls and remove obsolete enablement and retention controls. This complete before/after example preserves an existing chunk-plan cache and checkpoint directory while choosing an output and debug root explicitly.

# Version 2 (no longer accepted)
version: 2
workspace:
  directory: /srv/notarius/state
  resume:
    enabled: true
  debug:
    enabled: true
  chunk_cache:
    directory: /srv/notarius/chunk-plans
    mode: auto
  diagnostics:
    retention: always
diagnostics:
  work_dir: /srv/notarius/inspection

# Version 3
version: 3
output:
  directory: /srv/notarius/output
cache:
  chunk_plans:
    directory: /srv/notarius/chunk-plans
    mode: auto
  checkpoints:
    enabled: true
    directory: /srv/notarius/state/checkpoints
debug:
  directory: /srv/notarius/debug

Run the migrated configuration with --resume when checkpoint reuse is wanted, and with --debug when a debug bundle is wanted. Enabled checkpoint recording occurs with or without --resume.

The removed fields are workspace.directory, workspace.resume.enabled, workspace.debug.enabled, workspace.chunk_cache.mode, workspace.chunk_cache.directory, workspace.diagnostics.enabled, workspace.diagnostics.retention, diagnostics.work_dir, and diagnostics.retention. The removed environment variables are NOTARIUS_WORKSPACE_DIR, NOTARIUS_WORKSPACE_DIAGNOSTICS_ENABLED, NOTARIUS_WORKSPACE_DIAGNOSTICS_RETENTION, NOTARIUS_WORKSPACE_RESUME_ENABLED, NOTARIUS_WORKSPACE_DEBUG_ENABLED, NOTARIUS_WORKSPACE_CHUNK_CACHE_MODE, NOTARIUS_WORKSPACE_CHUNK_CACHE_DIR, NOTARIUS_WORK_DIR, and NOTARIUS_DIAGNOSTICS_RETENTION. The chunk-cache variables are replaced by NOTARIUS_CACHE_CHUNK_PLANS_MODE and NOTARIUS_CACHE_CHUNK_PLANS_DIR; the former shared directory has no direct replacement.

Validation

Configuration validation checks:

supported config version and known YAML fields;
mutually exclusive scriptorium.profile_dir and scriptorium.profile_file;
non-empty, non-duplicated IDs after trimming;
positive global LLM concurrency;
supported stage-worker keys and an effective extract worker count in the inclusive range 1..concurrency.total_llm;
non-empty output and debug directories;
a supported chunk-cache mode, Boolean checkpoint enablement, and state-surface directories without NUL bytes;
stale removed fields such as llm_profiles.

Pipeline resolution additionally checks:

the pipeline ID exists;
at least one artifact lane is declared and selected in each explicit step;
artifacts and steps are mutually exclusive, explicit steps are non-empty, and step IDs are unique after trimming;
lanes selected through the CLI exist in the resolved pipeline;
lane IDs are globally unique across ordered steps;
required module keys are present;
module keys are registered for the expected slot;
module capability requirements are satisfied;
non-empty validator overrides reference registered validator keys;
deterministic validators do not set llm_profile;
LLM-backed validators with explicit llm_profile values reference configured Scriptorium profile IDs;
bound reference slots are declared by selected chunk, extractor, merger, or normalizer targets;
generated references identify one lane in an earlier step, use a declared compatible artifact kind, and do not conflict with external or target-local generated bindings;
required reference slots are bound for selected targets.

28 KiB Raw Blame History