lpgcast.cast

Cast episodes into a governed schema: the whole pipeline.

Per episode: select a slice → one extraction call → validate → at most two typed repairs → quarantine what still fails → resolve identities and assemble. No critic pass: a valid extraction is never judged a second time.

An episode that produces nothing is quarantined and the run goes on; so is one whose request fails at the provider (provider_error), whose reply stops at the output limit (truncated) or that the model refuses (refused). A fault in the deployment — no key, a refused configuration, a cache miss while replaying — stops the run, because every other episode would fail the same way.

Attributes

EPISODE_FAULTS = (LLMReplyError, LLMRequestTimeoutError, APIError) module-attribute

EpisodeStatus = Literal['ok', 'partial', 'failed'] module-attribute

MAX_REPAIRS = 2 module-attribute

OUTPUT_FORMAT_VERSION = 2 module-attribute

Classes

CastResult

Bases: BaseModel

The output of a cast.

Source code in lpgcast/cast.py

class CastResult(BaseModel):
    """The output of a cast."""

    model_config = ConfigDict(arbitrary_types_allowed=True)

    output: GraFloOutput = Field(
        description="The full schema and every episode's graph."
    )
    pending: list[PendingEdge] = Field(default_factory=list)
    groundings: list[Grounding] = Field(default_factory=list)
    episodes: list[EpisodeResult] = Field(default_factory=list)
    manifest: RunManifest

Attributes

episodes = Field(default_factory=list) class-attribute instance-attribute

groundings = Field(default_factory=list) class-attribute instance-attribute

manifest instance-attribute

model_config = ConfigDict(arbitrary_types_allowed=True) class-attribute instance-attribute

output = Field(description="The full schema and every episode's graph.") class-attribute instance-attribute

pending = Field(default_factory=list) class-attribute instance-attribute

EpisodeOutcome

Bases: BaseModel

One episode's result with the intermediate steps that produced it.

Source code in lpgcast/cast.py

class EpisodeOutcome(BaseModel):
    """One episode's result with the intermediate steps that produced it."""

    model_config = ConfigDict(arbitrary_types_allowed=True)

    result: EpisodeResult
    selection: Selection | None = Field(
        default=None, description="The slice; None when selection failed."
    )
    first: Validation | None = Field(
        default=None, description="The first reply, checked; None when there was none."
    )
    validation: Validation | None = Field(
        default=None, description="The attempt kept, after repairs."
    )
    resolution: Resolution | None = None
    quarantine: list[QuarantineRecord] = Field(
        default_factory=list, description="What the episode left out."
    )

Attributes

first = Field(default=None, description='The first reply, checked; None when there was none.') class-attribute instance-attribute

model_config = ConfigDict(arbitrary_types_allowed=True) class-attribute instance-attribute

quarantine = Field(default_factory=list, description='What the episode left out.') class-attribute instance-attribute

resolution = None class-attribute instance-attribute

result instance-attribute

selection = Field(default=None, description='The slice; None when selection failed.') class-attribute instance-attribute

validation = Field(default=None, description='The attempt kept, after repairs.') class-attribute instance-attribute

EpisodeResult

Bases: BaseModel

What happened to one episode. Free of timing and cache facts, so a replay reproduces it byte for byte.

Source code in lpgcast/cast.py

class EpisodeResult(BaseModel):
    """What happened to one episode. Free of timing and cache facts, so a
    replay reproduces it byte for byte."""

    source_id: str
    status: EpisodeStatus = Field(
        description="``ok``: nothing left out; ``partial``: some items quarantined; "
        "``failed``: nothing written."
    )
    failure: str | None = Field(
        default=None, description="Why a failed episode failed."
    )
    selection: Literal["whole", "slice"] | None = None
    retrieval: RetrievalMetrics | None = None
    requests: int = Field(
        default=0, description="Model requests, re-asks and repairs included."
    )
    repairs: int = Field(default=0, description="Typed repair rounds.")
    repair_failure: str | None = Field(
        default=None,
        description="Why a repair request failed; the best earlier attempt is kept.",
    )
    vertices: int = 0
    edges: int = 0
    pending: int = Field(
        default=0, description="Edges awaiting a secondary-identity lookup."
    )
    quarantined: int = 0
    unmapped: int = 0
    fixes: list[Fix] = Field(default_factory=list)

Attributes

edges = 0 class-attribute instance-attribute

failure = Field(default=None, description='Why a failed episode failed.') class-attribute instance-attribute

fixes = Field(default_factory=list) class-attribute instance-attribute

pending = Field(default=0, description='Edges awaiting a secondary-identity lookup.') class-attribute instance-attribute

quarantined = 0 class-attribute instance-attribute

repair_failure = Field(default=None, description='Why a repair request failed; the best earlier attempt is kept.') class-attribute instance-attribute

repairs = Field(default=0, description='Typed repair rounds.') class-attribute instance-attribute

requests = Field(default=0, description='Model requests, re-asks and repairs included.') class-attribute instance-attribute

retrieval = None class-attribute instance-attribute

selection = None class-attribute instance-attribute

source_id instance-attribute

status = Field(description='``ok``: nothing left out; ``partial``: some items quarantined; ``failed``: nothing written.') class-attribute instance-attribute

unmapped = 0 class-attribute instance-attribute

vertices = 0 class-attribute instance-attribute

LPGCaster

Casts episodes into one governed schema.

Parameters:

Name	Type	Description	Default
schema	Schema	The governed schema; every write conforms to it.	required
settings	LPGCastSettings | None	Defaults to the environment.	None
client	LLMClient | None	Defaults to one built from settings.	None
sink	QuarantineSink | None	Where quarantined records go. Defaults to memory (:attr:sink then holds them).	None
selection	SelectionConfig | None	Selection tuning; its budget defaults to the settings'.	None
max_concurrent_episodes	int	Episodes in flight at once.	8

Source code in lpgcast/cast.py

class LPGCaster:
    """Casts episodes into one governed schema.

    Args:
        schema: The governed schema; every write conforms to it.
        settings: Defaults to the environment.
        client: Defaults to one built from *settings*.
        sink: Where quarantined records go. Defaults to memory
            (:attr:`sink` then holds them).
        selection: Selection tuning; its budget defaults to the settings'.
        max_concurrent_episodes: Episodes in flight at once.
    """

    def __init__(
        self,
        schema: Schema,
        *,
        settings: LPGCastSettings | None = None,
        client: LLMClient | None = None,
        sink: QuarantineSink | None = None,
        selection: SelectionConfig | None = None,
        max_concurrent_episodes: int = 8,
    ) -> None:
        self.schema = schema
        self.settings = settings or LPGCastSettings()
        self.client = client or LLMClient(
            self.settings.llm, LLMCache(self.settings.cache)
        )
        self.sink: QuarantineSink = sink if sink is not None else MemoryQuarantineSink()
        self.selection = selection or SelectionConfig(
            max_tokens=self.settings.slice_max_tokens,
            max_elements=self.settings.slice_max_elements,
        )
        self.max_concurrent_episodes = max(1, max_concurrent_episodes)

    def _fail(
        self,
        episode: Episode,
        failure: str,
        message: str,
        *,
        raw: str | None = None,
        sliced: Selection | None = None,
        **known: Any,
    ) -> EpisodeOutcome:
        record = QuarantineRecord(
            record="episode",
            source_id=episode.source_id,
            kind=failure,
            message=message,
            raw_output=raw,
            attempts=known.get("requests", 0),
            schema_ref=schema_ref(self.schema),
        )
        return EpisodeOutcome(
            result=EpisodeResult(
                source_id=episode.source_id, status="failed", failure=failure, **known
            ),
            selection=sliced,
            quarantine=[record],
        )

    async def cast_episode(self, episode: Episode) -> EpisodeOutcome:
        """Cast one episode and write what it left out to the sink."""
        outcome = await self._cast_episode(episode)
        await self.sink.write_records(outcome.quarantine)
        return outcome

    async def _cast_episode(self, episode: Episode) -> EpisodeOutcome:
        try:
            selection = select(self.schema, episode.text, self.selection)
        except SelectionError as error:
            return self._fail(
                episode,
                f"selection_{error.reason}",
                str(error),
                retrieval=error.metrics,
            )
        slice_ = selection.slice
        known: dict[str, Any] = {
            "selection": selection.mode,
            "retrieval": selection.metrics,
        }
        try:
            extraction = await extract(
                self.client,
                slice_,
                episode,
                user_instruction=self.settings.user_instruction,
                max_episode_tokens=self.settings.max_episode_tokens,
            )
        except EpisodeTooLongError as error:
            return self._fail(
                episode, "too_long", str(error), sliced=selection, **known
            )
        except WireShapeError as error:
            return self._fail(
                episode,
                "wire_shape",
                str(error),
                raw=error.content,
                sliced=selection,
                **known,
            )
        except LLMJsonParseError as error:
            return self._fail(
                episode,
                "unparseable",
                str(error),
                raw=error.doc,
                sliced=selection,
                **known,
            )
        except EPISODE_FAULTS as error:
            return self._fail(
                episode,
                _fault_kind(error),
                str(error),
                raw=error.content if isinstance(error, LLMReplyError) else None,
                sliced=selection,
                requests=1,
                **known,
            )

        requests = extraction.completion.attempts
        messages = build_messages(
            slice_, episode, user_instruction=self.settings.user_instruction
        )
        shape = instance_json_schema(slice_)
        current = first = validate(
            extraction.wire, slice_, episode.text, episode.context
        )
        current_reply = extraction.completion.content
        best, best_reply = current, current_reply
        repairs = 0
        repair_failure: str | None = None
        while current.repairable and repairs < MAX_REPAIRS:
            repairs += 1
            asked = _findings_key(current.repairable)
            try:
                completion = await self.client.acomplete_json(
                    repair_messages(messages, current_reply, current.repairable),
                    shape,
                    schema_name=INSTANCE_SCHEMA_NAME,
                )
                requests += completion.attempts
                wire = WireExtraction.model_validate(completion.data)
            except LLMJsonParseError:
                requests += 1
                repair_failure = "unparseable"
                break
            except ValidationError:
                repair_failure = "wire_shape"
                break
            except EPISODE_FAULTS as error:
                requests += 1
                repair_failure = _fault_kind(error)
                break
            current = validate(wire, slice_, episode.text, episode.context)
            current_reply = completion.content
            if current.kept >= best.kept:
                best, best_reply = current, current_reply
            # The same findings again: the slice does not let the model fix them.
            if _findings_key(current.repairable) == asked:
                break

        known.update(requests=requests, repairs=repairs, repair_failure=repair_failure)
        quarantine = self._records(episode, best, best_reply, requests)
        resolution = resolve(
            best,
            slice_,
            episode,
            extractor=_extractor(),
            model=self.client.settings.model,
        )
        container = resolution.container
        kept = best.kept
        status: EpisodeStatus = (
            "ok" if not best.findings else "partial" if kept else "failed"
        )
        return EpisodeOutcome(
            result=EpisodeResult(
                source_id=episode.source_id,
                status=status,
                failure="nothing_valid" if status == "failed" else None,
                vertices=sum(len(docs) for docs in container.vertices.values()),
                edges=sum(len(triples) for triples in container.edges.values()),
                pending=len(resolution.pending),
                quarantined=len(best.findings),
                unmapped=len(best.unmapped),
                fixes=best.fixes,
                **known,
            ),
            selection=selection,
            first=first,
            validation=best,
            resolution=resolution,
            quarantine=quarantine,
        )

    def _records(
        self, episode: Episode, validation: Validation, reply: str, requests: int
    ) -> list[QuarantineRecord]:
        ref = schema_ref(self.schema)
        records = [
            QuarantineRecord(
                record="finding",
                source_id=episode.source_id,
                kind=finding.kind,
                detail=finding.detail,
                path=finding.path,
                message=finding.message,
                quote=finding.fragment.get("quote"),
                fragment=finding.fragment,
                attempts=requests,
                raw_output=reply,
                schema_ref=ref,
            )
            for finding in validation.findings
        ]
        records += [
            QuarantineRecord(
                record="unmapped",
                source_id=episode.source_id,
                kind="unmapped",
                message=item.note,
                quote=item.quote,
                char_start=item.span.start if item.span else None,
                char_end=item.span.end if item.span else None,
                schema_ref=ref,
            )
            for item in validation.unmapped
        ]
        return records

    async def cast(self, episodes: Iterable[Episode]) -> CastResult:
        """Cast *episodes*; results and quarantine records keep their order.

        Raises:
            LLMConfigurationError: the deployment is misconfigured.
            CacheMissError: a miss while the cache requires hits.
        """
        items = list(episodes)
        gate = asyncio.Semaphore(self.max_concurrent_episodes)

        async def _one(episode: Episode) -> EpisodeOutcome:
            async with gate:
                return await self._cast_episode(episode)

        with track_usage() as usage:
            try:
                async with asyncio.TaskGroup() as group:
                    tasks = [group.create_task(_one(episode)) for episode in items]
            except ExceptionGroup as errors:
                raise errors.exceptions[0] from errors
        outcomes = [task.result() for task in tasks]
        for outcome in outcomes:
            await self.sink.write_records(outcome.quarantine)
        return self._assemble(outcomes, usage)

    def _assemble(
        self, outcomes: list[EpisodeOutcome], usage: UsageTracker
    ) -> CastResult:
        resolutions = [o.resolution for o in outcomes if o.resolution is not None]
        container = merge_containers([r.container for r in resolutions])
        results = [o.result for o in outcomes]
        status = {name: 0 for name in ("ok", "partial", "failed")}
        for result in results:
            status[result.status] += 1
        manifest = RunManifest(
            lpgcast_version=__version__,
            graflo_version=version("graflo"),
            schema_ref=schema_ref(self.schema),
            model={
                key: value
                for key, value in cache_key_config(self.client.settings).items()
                if key != "cache_format_version"
            },
            max_episode_tokens=self.settings.max_episode_tokens,
            slice_max_tokens=self.selection.max_tokens,
            slice_max_elements=self.selection.max_elements,
            user_instruction=bool(self.settings.user_instruction),
            episodes=len(results),
            status=status,
            vertices=sum(r.vertices for r in results),
            edges=sum(r.edges for r in results),
            pending=sum(r.pending for r in results),
            quarantined=sum(r.quarantined for r in results),
            unmapped=sum(r.unmapped for r in results),
            usage=usage,
        )
        return CastResult(
            output=GraFloOutput.model_validate(
                {"schema": self.schema, "data": container}
            ),
            pending=[edge for r in resolutions for edge in r.pending],
            groundings=[g for r in resolutions for g in r.groundings],
            episodes=results,
            manifest=manifest,
        )

Attributes

client = client or LLMClient(self.settings.llm, LLMCache(self.settings.cache)) instance-attribute

max_concurrent_episodes = max(1, max_concurrent_episodes) instance-attribute

schema = schema instance-attribute

selection = selection or SelectionConfig(max_tokens=self.settings.slice_max_tokens, max_elements=self.settings.slice_max_elements) instance-attribute

settings = settings or LPGCastSettings() instance-attribute

sink = sink if sink is not None else MemoryQuarantineSink() instance-attribute

Methods:

__init__(schema, *, settings=None, client=None, sink=None, selection=None, max_concurrent_episodes=8)

Source code in lpgcast/cast.py

def __init__(
    self,
    schema: Schema,
    *,
    settings: LPGCastSettings | None = None,
    client: LLMClient | None = None,
    sink: QuarantineSink | None = None,
    selection: SelectionConfig | None = None,
    max_concurrent_episodes: int = 8,
) -> None:
    self.schema = schema
    self.settings = settings or LPGCastSettings()
    self.client = client or LLMClient(
        self.settings.llm, LLMCache(self.settings.cache)
    )
    self.sink: QuarantineSink = sink if sink is not None else MemoryQuarantineSink()
    self.selection = selection or SelectionConfig(
        max_tokens=self.settings.slice_max_tokens,
        max_elements=self.settings.slice_max_elements,
    )
    self.max_concurrent_episodes = max(1, max_concurrent_episodes)

cast(episodes) async

Cast episodes; results and quarantine records keep their order.

Raises:

Type	Description
LLMConfigurationError	the deployment is misconfigured.
CacheMissError	a miss while the cache requires hits.

Source code in lpgcast/cast.py

async def cast(self, episodes: Iterable[Episode]) -> CastResult:
    """Cast *episodes*; results and quarantine records keep their order.

    Raises:
        LLMConfigurationError: the deployment is misconfigured.
        CacheMissError: a miss while the cache requires hits.
    """
    items = list(episodes)
    gate = asyncio.Semaphore(self.max_concurrent_episodes)

    async def _one(episode: Episode) -> EpisodeOutcome:
        async with gate:
            return await self._cast_episode(episode)

    with track_usage() as usage:
        try:
            async with asyncio.TaskGroup() as group:
                tasks = [group.create_task(_one(episode)) for episode in items]
        except ExceptionGroup as errors:
            raise errors.exceptions[0] from errors
    outcomes = [task.result() for task in tasks]
    for outcome in outcomes:
        await self.sink.write_records(outcome.quarantine)
    return self._assemble(outcomes, usage)

cast_episode(episode) async

Cast one episode and write what it left out to the sink.

Source code in lpgcast/cast.py

async def cast_episode(self, episode: Episode) -> EpisodeOutcome:
    """Cast one episode and write what it left out to the sink."""
    outcome = await self._cast_episode(episode)
    await self.sink.write_records(outcome.quarantine)
    return outcome

RunManifest

Bases: BaseModel

How a cast was produced.

Source code in lpgcast/cast.py

class RunManifest(BaseModel):
    """How a cast was produced."""

    lpgcast_version: str
    graflo_version: str
    output_format_version: int = OUTPUT_FORMAT_VERSION
    schema_ref: str
    model: dict[str, Any] = Field(
        description="Every model setting that reaches the provider."
    )
    max_episode_tokens: int
    slice_max_tokens: int | None
    slice_max_elements: int | None
    user_instruction: bool
    episodes: int
    status: dict[str, int]
    vertices: int
    edges: int
    pending: int
    quarantined: int
    unmapped: int
    usage: UsageTracker

Attributes

edges instance-attribute

episodes instance-attribute

graflo_version instance-attribute

lpgcast_version instance-attribute

max_episode_tokens instance-attribute

model = Field(description='Every model setting that reaches the provider.') class-attribute instance-attribute

output_format_version = OUTPUT_FORMAT_VERSION class-attribute instance-attribute

pending instance-attribute

quarantined instance-attribute

schema_ref instance-attribute

slice_max_elements instance-attribute

slice_max_tokens instance-attribute

status instance-attribute

unmapped instance-attribute

usage instance-attribute

user_instruction instance-attribute

vertices instance-attribute

Functions: