core/generate/stream.py

391 lines
13 KiB
Python

"""
Generation loop — token streaming from the versor manifold.
Every token: nearest non-current word to current F via CGA inner product.
Every step: F <- versor_apply(V, F) where V = word_transition_rotor(A, B).
Architectural boundaries enforced here:
- VocabManifold owns manifold points only (get_versor_at, nearest).
- algebra.rotor.word_transition_rotor constructs the transition operator.
- Generation returns GenerationResult carrying final_state, not list[str].
- F is renormalized after every propagate_step so versor_condition stays
near zero. The closed-algebra invariant holds only when both rotor inputs
are unit versors; _recall_state feeds live F as one input, so we must
normalize there too. See ADR note below.
ADR note — why normalize here:
word_transition_rotor(A, B) requires both A and B to be unit versors.
Inside the main loop A is always vocab.get_versor_at(node) (safe).
Inside _recall_state A is current.F which drifts under repeated
sandwiching. Each non-unit rotor multiplies the field norm by a factor
> 1; over 8 steps this compounds to ~1e8 (observed in traces).
Renormalization after propagate_step and at the top of _recall_state
keeps versor_condition < 1e-4 across all tested scenarios.
No confidence gates. No IDK fallback. No attractor clamping.
"""
from __future__ import annotations
from collections import deque
import numpy as np
from field.state import FieldState
from field.propagate import propagate_step
from algebra.rotor import word_transition_rotor
from algebra.versor import unitize_versor
from generate.attention import AttentionOperator
from generate.result import GenerationResult
from generate.salience import SalienceOperator
_RECENT_WINDOW = 3
_STOP_TOKENS = frozenset({"it", "to", "word"})
def _renorm(state: FieldState) -> FieldState:
"""
Return state with F renormalized to unit versor norm.
This is called after every propagate_step to keep F on the manifold.
If F is already unit (norm within 1e-9 of 1.0) the copy is skipped and
the original state is returned unchanged.
"""
norm = float(np.linalg.norm(state.F))
if norm < 1e-12:
return state
if abs(norm - 1.0) < 1e-9:
return state
return FieldState(
F=state.F / norm,
node=state.node,
step=state.step,
holonomy=state.holonomy,
energy=state.energy,
valence=state.valence,
)
def _articulate(vocab, word: str) -> str:
"""
Recover the emitted surface through MorphologyEntry when available.
The manifold walk selects a vocabulary point. Articulation then returns
the structured surface carried by that point, preserving script and
inflection without introducing a corrective pass.
"""
morphology_for_word = getattr(vocab, "morphology_for_word", None)
if morphology_for_word is None:
return word
morphology = morphology_for_word(word)
return morphology.surface if morphology is not None else word
def _nearest_next(
vocab,
F_voiced,
current_node: int,
recent_nodes: tuple[int, ...] = (),
stop_nodes: frozenset[int] = frozenset(),
candidate_indices: np.ndarray | None = None,
) -> tuple[str, int]:
"""
Select the nearest vocabulary point while avoiding short loops.
Allowing the current node to win makes V = transition(A, A), which is an
identity-like transition and can stall generation forever on one token.
Recent-node exclusion reduces two- and three-token attractor cycles.
Stop-node exclusion keeps function-word wells from dominating when more
informative neighbors are available.
If attention/language filtering leaves only the current node available,
the final fallback deliberately permits that singleton candidate instead
of crashing. That keeps inhibition fail-closed to the attended region.
"""
if len(vocab) <= 1:
return vocab.nearest(F_voiced, candidate_indices=candidate_indices)
recent = set(recent_nodes)
stop = set(stop_nodes)
fallback_orders = (
recent | stop,
stop,
recent,
set(),
)
for extra in fallback_orders:
try:
return _nearest_with_optional_candidates(
vocab,
F_voiced,
current_node,
extra,
candidate_indices,
)
except ValueError:
continue
return _nearest_with_optional_candidates(
vocab,
F_voiced,
-1,
set(),
candidate_indices,
)
def _nearest_with_optional_candidates(
vocab,
F_voiced,
current_node: int,
exclude_indices: set[int],
candidate_indices: np.ndarray | None,
) -> tuple[str, int]:
try:
return vocab.nearest(
F_voiced,
exclude_idx=current_node,
exclude_indices=exclude_indices,
candidate_indices=candidate_indices,
)
except TypeError:
if candidate_indices is not None:
raise
return vocab.nearest(
F_voiced,
exclude_idx=current_node,
exclude_indices=exclude_indices,
)
def _voiced_state(state: FieldState, persona) -> FieldState:
"""Compose the session persona motor into the live field path."""
return _renorm(FieldState(
F=persona.apply(state.F),
node=state.node,
step=state.step,
holonomy=state.holonomy,
energy=state.energy,
valence=state.valence,
))
def _recall_state(state: FieldState, vault, top_k: int) -> FieldState:
"""
Feed exact vault recall back into the field as sequential operators.
Recall returns stored versors ranked by the vault's exact metric. Each hit
is treated as an additional operator in the propagation path.
IMPORTANT: current.F must be unit before passing to word_transition_rotor
as input A. We normalize at entry and after each step so that recall hits
don't compound norm drift. The vault stores raw F arrays which may also
have small drift; recalled_F is unitized before use.
"""
if vault is None or top_k <= 0:
return state
current = _renorm(state)
for hit in vault.recall(current.F, top_k=top_k):
recalled_F = np.asarray(hit["versor"], dtype=np.float64)
r_norm = float(np.linalg.norm(recalled_F))
if r_norm > 1e-12:
recalled_F = recalled_F / r_norm
V = word_transition_rotor(current.F, recalled_F)
current = _renorm(propagate_step(current, V))
current = FieldState(
F=current.F,
node=state.node,
step=current.step,
holonomy=state.holonomy,
energy=state.energy,
valence=state.valence,
)
return current
def _candidate_indices_for_language(vocab, output_lang: str | None) -> np.ndarray | None:
if output_lang is None:
return None
indices_for_language = getattr(vocab, "indices_for_language", None)
if indices_for_language is None:
return None
indices = indices_for_language(output_lang)
if len(indices) == 0:
raise ValueError(f"No generation candidates for output language {output_lang!r}.")
return indices
def _intersect_candidates(a: np.ndarray | None, b: np.ndarray | None) -> np.ndarray | None:
if a is None:
return b
if b is None:
return a
if len(a) == 0 or len(b) == 0:
return np.asarray([], dtype=np.int64)
b_set = {int(idx) for idx in b}
return np.asarray([int(idx) for idx in a if int(idx) in b_set], dtype=np.int64)
def _attention_candidates(
state: FieldState,
vocab,
use_salience: bool,
salience_top_k: int,
inhibition_threshold: float,
) -> tuple[np.ndarray | None, int | None, int | None]:
if not use_salience:
return None, None, None
salience = SalienceOperator().compute(state, vocab, top_k=salience_top_k)
attention = AttentionOperator(inhibition_threshold).plan(salience, vocab)
return attention.allowed_indices, salience.budget, len(attention.allowed_indices)
def generate(
state: FieldState,
vocab,
persona,
max_tokens: int = 128,
record_trajectory: bool = False,
vault=None,
recall_top_k: int = 3,
output_lang: str | None = None,
allow_cross_language_generation: bool = True,
use_salience: bool = False,
salience_top_k: int = 16,
inhibition_threshold: float = 0.3,
) -> GenerationResult:
"""
Generate a token sequence from an initial FieldState.
Loop:
1. Compose the persistent persona motor into the current field
2. Propagate exact vault recall hits into the current field
3. Find nearest non-current vocab node via CGA inner product
4. Emit token
5. Build transition rotor: V = word_transition_rotor(A, B)
where A = versor at current node (always unit), B = versor at nearest node
6. Propagate: F <- versor_apply(V, F)
7. Renormalize F to keep it on the manifold (versor_condition < 1e-4)
8. Advance node pointer
Returns:
GenerationResult with tokens, final_state, optional trajectory,
and salience telemetry when attention is enabled.
"""
tokens = []
trajectory = [] if record_trajectory else None
current = _renorm(state)
recent_nodes = deque([state.node], maxlen=_RECENT_WINDOW)
language_candidates = None if allow_cross_language_generation else _candidate_indices_for_language(vocab, output_lang)
salience_candidates, salience_budget, candidates_used = _attention_candidates(
state,
vocab,
use_salience=use_salience,
salience_top_k=salience_top_k,
inhibition_threshold=inhibition_threshold,
)
candidate_indices = _intersect_candidates(language_candidates, salience_candidates)
if candidate_indices is not None and len(candidate_indices) == 0:
candidate_indices = salience_candidates if salience_candidates is not None else language_candidates
candidates_used = None if candidate_indices is None else len(candidate_indices)
stop_nodes = frozenset(
vocab.index_of(token)
for token in _STOP_TOKENS
if token in {vocab.get_word_at(i) for i in range(len(vocab))}
)
token_budget = min(max_tokens, int(candidates_used)) if candidates_used is not None else max_tokens
for _ in range(token_budget):
current = _recall_state(_voiced_state(current, persona), vault, recall_top_k)
word, word_idx = _nearest_next(
vocab,
current.F,
current.node,
recent_nodes=tuple(recent_nodes),
stop_nodes=stop_nodes,
candidate_indices=candidate_indices,
)
tokens.append(_articulate(vocab, word))
if record_trajectory:
trajectory.append(current)
A = vocab.get_versor_at(current.node)
B = vocab.get_versor_at(word_idx)
V = word_transition_rotor(A, B)
current = _renorm(propagate_step(current, V))
current = FieldState(
F=current.F,
node=word_idx,
step=current.step,
holonomy=current.holonomy,
energy=current.energy,
valence=current.valence,
)
recent_nodes.append(word_idx)
return GenerationResult(
tokens=tokens,
final_state=current,
trajectory=trajectory,
salience_top_k=salience_budget,
candidates_used=candidates_used,
)
async def agenerate(
state: FieldState,
vocab,
persona,
max_tokens: int = 128,
vault=None,
recall_top_k: int = 3,
):
"""
Async streaming version — yields one token at a time.
Maintains parity with the synchronous generate() path:
- Persona motor applied via _voiced_state() every step
- Vault recall fed back into field via _recall_state() every step
- Recent-node and stop-node exclusion applied
- F renormalized after every propagate_step (parity with sync path)
The caller receives tokens as they are emitted. For the full
GenerationResult (final_state, trajectory), use the synchronous
generate() path or wrap this generator in an async collector.
Yields: str (one token per iteration)
"""
current = _renorm(state)
recent_nodes = deque([state.node], maxlen=_RECENT_WINDOW)
stop_nodes = frozenset(
vocab.index_of(token)
for token in _STOP_TOKENS
if token in {vocab.get_word_at(i) for i in range(len(vocab))}
)
for _ in range(max_tokens):
current = _recall_state(_voiced_state(current, persona), vault, recall_top_k)
word, word_idx = _nearest_next(
vocab,
current.F,
current.node,
recent_nodes=tuple(recent_nodes),
stop_nodes=stop_nodes,
)
yield _articulate(vocab, word)
A = vocab.get_versor_at(current.node)
B = vocab.get_versor_at(word_idx)
V = word_transition_rotor(A, B)
current = _renorm(propagate_step(current, V))
current = FieldState(
F=current.F,
node=word_idx,
step=current.step,
holonomy=current.holonomy,
energy=current.energy,
valence=current.valence,
)
recent_nodes.append(word_idx)