feat(resolver): implement the deterministic Illustrious Resolver (#12) (#56)

Add compiler/resolver/illustrious_resolver.py: resolve_scene(scene,
knowledge_base, config, logger=None) -> CompilerResult, per MASTER_SPEC §17. No
LLM; no concept or tag is invented.

- normalize_concept (§17.3): NFKC + lowercase, unifying underscores/spaces so
  natural concepts, snake_case ids, and spaced aliases share one lookup key. The
  resolver builds its own normalized index over the Knowledge Base.
- Per-concept pipeline in scene discovery order (§17.8): normalize -> direct
  alias resolution (no chains; skipped when allow_aliases is false) -> canonical
  lookup -> recursive expansion -> tag generation.
- Expansion (§17.5): own tags first, then depth-first in list order per expand
  target, inserted immediately after; capped by max_expansion_depth; skipped when
  expansion_enabled is false; defensive path-based cycle detection -> SC0003.
- Deduplication after expansion keeps the first occurrence (SC0007).
- Unknown concept -> SC0001 warning (continue); deprecated -> SC0006 (continue);
  invalid category -> SC0008 (stop); circular expansion -> SC0003 (stop).
- Every ResolvedTag keeps its entry's category and full traceability
  (source_concept = originating scene concept, knowledge_base_entry = generating
  entry). Verbose logging of each decision.

Tests (no Ollama): normalization, alias resolution + disable, expansion order,
nested/depth-capped/disabled expansion, traceability, dedup, unknown/deprecated,
circular/invalid-category stops, category preservation, and determinism.

Co-authored-by: kunail0804 <kunail0804@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Crespin Alexandre
2026-07-07 19:58:12 +02:00
committed by GitHub
co-authored by kunail0804 Claude Opus 4.8
parent 3f0358bfa9
commit 76d440fbaf
2 changed files with 479 additions and 0 deletions
+250
View File
@@ -0,0 +1,250 @@
"""The Illustrious Resolver (MASTER_SPEC §17).
Deterministically converts the semantic concepts of a validated Scene into
Illustrious Resolved Tags using the Knowledge Base. No language model is involved
and no concept or tag is ever invented.
Per-concept pipeline (§17.2): normalize -> alias resolution (direct only) ->
canonical lookup -> expansion -> tag generation. Concepts are processed in scene
discovery order (§17.8) and every generated tag keeps the category of its
Knowledge Base Entry.
"""
from __future__ import annotations
import unicodedata
from collections.abc import Iterator
from compiler.common.categories import is_valid_category
from compiler.common.config import Config
from compiler.common.knowledge_base import KnowledgeBase, KnowledgeBaseEntry
from compiler.common.log import StructuredLogger
from compiler.common.message_codes import message
from compiler.common.result import CompilerResult, Message
from schemas.models import ResolvedTag, Scene
_CHARACTER_CONCEPT_FIELDS = (
"identity",
"appearance",
"clothing",
"accessories",
"pose",
"expression",
"actions",
)
def normalize_concept(name: str) -> str:
"""Normalize a concept for lookup (§17.3).
Applies Unicode (NFKC) normalization and lowercasing, and unifies underscores
with spaces before collapsing whitespace, so natural-language concepts
("holding hands"), snake_case ids ("holding_hands"), and spaced aliases all
map to the same key.
"""
text = unicodedata.normalize("NFKC", name).replace("_", " ").lower()
return " ".join(text.split())
def resolve_scene(
scene: Scene,
knowledge_base: KnowledgeBase,
config: Config,
logger: StructuredLogger | None = None,
) -> CompilerResult:
"""Resolve a validated Scene into an ordered tuple of Resolved Tags.
Returns:
A CompilerResult whose ``data`` is the tuple of :class:`ResolvedTag` on
success, or ``None`` with error messages when an error condition
(circular expansion SC0003, invalid category SC0008) stops resolution.
"""
index = _build_normalized_index(knowledge_base)
resolver = _ConceptResolver(knowledge_base, index, config, logger)
tags: list[ResolvedTag] = []
warnings: list[Message] = []
errors: list[Message] = []
for concept_name in _iter_scene_concepts(scene):
resolved = resolver.resolve(concept_name)
tags.extend(resolved.tags)
warnings.extend(resolved.warnings)
errors.extend(resolved.errors)
if errors:
result = CompilerResult()
for error in errors:
result = result.add_error(error)
return result
deduped, dedup_warnings = _deduplicate(tags)
warnings.extend(dedup_warnings)
result = CompilerResult(data=tuple(deduped))
for warning in warnings:
result = result.add_warning(warning)
if logger is not None:
logger.basic("scene_resolved", tags=len(deduped), warnings=len(warnings))
return result
def _build_normalized_index(
knowledge_base: KnowledgeBase,
) -> dict[str, KnowledgeBaseEntry]:
"""Map every normalized canonical id and alias to its entry (aliases direct)."""
index: dict[str, KnowledgeBaseEntry] = {}
for entry in knowledge_base.by_id.values():
index[normalize_concept(entry.id)] = entry
for alias in entry.aliases:
index[normalize_concept(alias)] = entry
return index
def _iter_scene_concepts(scene: Scene) -> Iterator[str]:
"""Yield concept names in scene discovery order (§17.8)."""
for character in scene.characters:
for field in _CHARACTER_CONCEPT_FIELDS:
for concept in getattr(character, field):
yield concept.name
for interaction in scene.interactions:
yield interaction.concept
for field in ("objects", "environment", "camera", "lighting"):
for concept in getattr(scene, field):
yield concept.name
class _Resolution:
"""The outcome of resolving one concept."""
__slots__ = ("tags", "warnings", "errors")
def __init__(self) -> None:
self.tags: list[ResolvedTag] = []
self.warnings: list[Message] = []
self.errors: list[Message] = []
class _ConceptResolver:
"""Resolves a single concept (with expansion) against a normalized index."""
def __init__(
self,
knowledge_base: KnowledgeBase,
index: dict[str, KnowledgeBaseEntry],
config: Config,
logger: StructuredLogger | None,
) -> None:
self._kb = knowledge_base
self._index = index
self._allow_aliases = config.resolver.allow_aliases
self._expansion_enabled = config.resolver.expansion_enabled
self._max_depth = config.resolver.max_expansion_depth
self._logger = logger
def resolve(self, concept_name: str) -> _Resolution:
outcome = _Resolution()
key = normalize_concept(concept_name)
entry = self._lookup(key)
if entry is None:
outcome.warnings.append(
message(
"SC0001",
f"Unknown concept '{concept_name}' has no Knowledge Base entry; ignored.",
concept=concept_name,
)
)
if self._logger is not None:
self._logger.verbose("concept_unknown", concept=concept_name, normalized=key)
return outcome
self._expand(entry, concept_name, depth=0, path=(), outcome=outcome)
return outcome
def _lookup(self, key: str) -> KnowledgeBaseEntry | None:
entry = self._index.get(key)
if entry is None:
return None
# A key that matches an alias but not the entry's own id is an alias hit.
if not self._allow_aliases and key != normalize_concept(entry.id):
return None
return entry
def _expand(
self,
entry: KnowledgeBaseEntry,
source_concept: str,
depth: int,
path: tuple[str, ...],
outcome: _Resolution,
) -> None:
if entry.id in path:
outcome.errors.append(
message(
"SC0003",
f"Circular expansion detected at '{entry.id}' (path: {' -> '.join(path)}).",
entry=entry.id,
)
)
return
if not is_valid_category(entry.category):
outcome.errors.append(
message(
"SC0008",
f"Entry '{entry.id}' has invalid category '{entry.category}'.",
entry=entry.id,
)
)
return
if entry.deprecated:
outcome.warnings.append(
message(
"SC0006",
f"Concept '{entry.id}' is deprecated; migration is encouraged.",
entry=entry.id,
)
)
for tag in entry.tags:
outcome.tags.append(
ResolvedTag(
tag=tag,
category=entry.category,
source_concept=source_concept,
knowledge_base_entry=entry.id,
)
)
if self._logger is not None:
self._logger.verbose(
"concept_resolved",
concept=source_concept,
canonical=entry.id,
tags=list(entry.tags),
)
if not self._expansion_enabled or depth >= self._max_depth:
return
for target_id in entry.expand:
target = self._kb.get(target_id)
if target is not None:
self._expand(target, source_concept, depth + 1, (*path, entry.id), outcome)
def _deduplicate(tags: list[ResolvedTag]) -> tuple[list[ResolvedTag], list[Message]]:
"""Remove duplicate tags after expansion, keeping the first occurrence (§17.7)."""
seen: set[str] = set()
kept: list[ResolvedTag] = []
warnings: list[Message] = []
for resolved in tags:
if resolved.tag in seen:
warnings.append(
message(
"SC0007",
f"Duplicate tag '{resolved.tag}' removed after expansion.",
tag=resolved.tag,
)
)
continue
seen.add(resolved.tag)
kept.append(resolved)
return kept, warnings
+229
View File
@@ -0,0 +1,229 @@
"""Tests for the Illustrious Resolver (issue #12, MASTER_SPEC §17). No Ollama."""
from __future__ import annotations
from compiler.common.config import Config
from compiler.common.knowledge_base import KnowledgeBase, KnowledgeBaseEntry
from compiler.resolver.illustrious_resolver import resolve_scene
from schemas.models import Scene
def kb(*entries: KnowledgeBaseEntry) -> KnowledgeBase:
return KnowledgeBase(list(entries))
def entry(id_, tags, category="hair", aliases=(), expand=(), deprecated=False):
return KnowledgeBaseEntry(
id=id_,
tags=tuple(tags),
category=category,
aliases=tuple(aliases),
expand=tuple(expand),
deprecated=deprecated,
)
def scene_with(identity=(), **fields) -> Scene:
character = {
"id": 0,
"identity": list(identity),
"appearance": list(fields.get("appearance", [])),
"clothing": list(fields.get("clothing", [])),
"accessories": [],
"pose": [],
"expression": [],
"actions": [],
}
return Scene.from_json(
{
"characters": [character],
"interactions": [],
"objects": list(fields.get("objects", [])),
"environment": list(fields.get("environment", [])),
"camera": [],
"lighting": [],
"metadata": {"schema_version": "1.0"},
}
)
def default_config() -> Config:
return Config()
def tags(result) -> list[str]:
return [t.tag for t in result.data]
def codes(messages) -> list[str]:
return [m.code for m in messages]
# --- normalization ---------------------------------------------------------
def test_normalization_matches_snake_case_id() -> None:
knowledge = kb(entry("long_hair", ["long hair"]))
for written in ["long hair", "Long Hair", " LONG_HAIR ", "long_hair"]:
result = resolve_scene(scene_with(identity=[written]), knowledge, default_config())
assert tags(result) == ["long hair"], written
# --- alias resolution ------------------------------------------------------
def test_alias_resolves_directly() -> None:
knowledge = kb(entry("female", ["1girl"], category="character", aliases=["girl", "woman"]))
result = resolve_scene(scene_with(identity=["girl"]), knowledge, default_config())
assert tags(result) == ["1girl"]
assert result.data[0].knowledge_base_entry == "female"
assert result.data[0].source_concept == "girl"
def test_alias_ignored_when_disabled() -> None:
knowledge = kb(entry("female", ["1girl"], category="character", aliases=["girl"]))
config = Config.from_json({"resolver": {"allow_aliases": False}})
result = resolve_scene(scene_with(identity=["girl"]), knowledge, config)
assert tags(result) == []
assert "SC0001" in codes(result.warnings)
# --- expansion order -------------------------------------------------------
def test_expansion_order_is_normative() -> None:
knowledge = kb(
entry(
"school_uniform", ["school uniform"], category="clothing", expand=["blazer", "skirt"]
),
entry("blazer", ["blazer"], category="clothing"),
entry("skirt", ["pleated skirt"], category="clothing"),
)
result = resolve_scene(scene_with(clothing=["school uniform"]), knowledge, default_config())
assert tags(result) == ["school uniform", "blazer", "pleated skirt"]
def test_nested_expansion_is_depth_first() -> None:
knowledge = kb(
entry("a", ["ta"], expand=["b"]),
entry("b", ["tb"], expand=["c"]),
entry("c", ["tc"]),
)
result = resolve_scene(scene_with(identity=["a"]), knowledge, default_config())
assert tags(result) == ["ta", "tb", "tc"]
def test_expansion_disabled_emits_only_own_tags() -> None:
knowledge = kb(
entry("a", ["ta"], expand=["b"]),
entry("b", ["tb"]),
)
config = Config.from_json({"resolver": {"expansion_enabled": False}})
result = resolve_scene(scene_with(identity=["a"]), knowledge, config)
assert tags(result) == ["ta"]
def test_max_expansion_depth_caps_recursion() -> None:
# depth 1 allows one expansion level (a -> b) but not the next (b -> c).
knowledge = kb(
entry("a", ["ta"], expand=["b"]),
entry("b", ["tb"], expand=["c"]),
entry("c", ["tc"]),
)
config = Config.from_json({"resolver": {"max_expansion_depth": 1}})
result = resolve_scene(scene_with(identity=["a"]), knowledge, config)
assert tags(result) == ["ta", "tb"]
def test_expanded_tag_traceability() -> None:
knowledge = kb(
entry("school_uniform", ["school uniform"], category="clothing", expand=["blazer"]),
entry("blazer", ["blazer"], category="clothing"),
)
result = resolve_scene(scene_with(clothing=["school uniform"]), knowledge, default_config())
blazer_tag = result.data[1]
assert blazer_tag.tag == "blazer"
assert blazer_tag.knowledge_base_entry == "blazer"
assert blazer_tag.source_concept == "school uniform" # originating scene concept
# --- deduplication ---------------------------------------------------------
def test_duplicate_tags_removed_keeping_first() -> None:
knowledge = kb(
entry("a", ["shared"]),
entry("b", ["shared"]),
)
result = resolve_scene(scene_with(identity=["a", "b"]), knowledge, default_config())
assert tags(result) == ["shared"]
assert "SC0007" in codes(result.warnings)
# --- unknown / deprecated --------------------------------------------------
def test_unknown_concept_warns_and_continues() -> None:
knowledge = kb(entry("female", ["1girl"], category="character"))
result = resolve_scene(scene_with(identity=["dragon", "female"]), knowledge, default_config())
assert tags(result) == ["1girl"]
assert "SC0001" in codes(result.warnings)
def test_deprecated_concept_warns_but_emits_tags() -> None:
knowledge = kb(entry("old_style", ["old style"], deprecated=True))
result = resolve_scene(scene_with(identity=["old_style"]), knowledge, default_config())
assert tags(result) == ["old style"]
assert "SC0006" in codes(result.warnings)
# --- error conditions stop -------------------------------------------------
def test_circular_expansion_stops_with_sc0003() -> None:
knowledge = kb(
entry("a", ["ta"], expand=["b"]),
entry("b", ["tb"], expand=["a"]),
)
result = resolve_scene(scene_with(identity=["a"]), knowledge, default_config())
assert not result.success
assert result.data is None
assert "SC0003" in codes(result.errors)
def test_invalid_category_stops_with_sc0008() -> None:
knowledge = kb(entry("a", ["ta"], category="not_a_category"))
result = resolve_scene(scene_with(identity=["a"]), knowledge, default_config())
assert not result.success
assert "SC0008" in codes(result.errors)
# --- category preservation & determinism -----------------------------------
def test_tag_keeps_entry_category() -> None:
knowledge = kb(entry("blue_eyes", ["blue eyes"], category="eyes"))
result = resolve_scene(scene_with(appearance=["blue_eyes"]), knowledge, default_config())
assert result.data[0].category == "eyes"
def test_discovery_order_preserved_across_scene() -> None:
knowledge = kb(
entry("female", ["1girl"], category="character"),
entry("forest", ["forest"], category="environment"),
)
scene = scene_with(identity=["female"], environment=["forest"])
result = resolve_scene(scene, knowledge, default_config())
assert tags(result) == ["1girl", "forest"]
def test_output_is_deterministic() -> None:
knowledge = kb(
entry("school_uniform", ["school uniform"], category="clothing", expand=["blazer"]),
entry("blazer", ["blazer"], category="clothing"),
)
scene = scene_with(clothing=["school uniform", "school uniform"])
a = resolve_scene(scene, knowledge, default_config())
b = resolve_scene(scene, knowledge, default_config())
assert tags(a) == tags(b)
assert codes(a.warnings) == codes(b.warnings)