Skip to main content
Version: 0.7.3

Primitives

Attribution​

@dataclass(frozen=True)
class Attribution

A word-level relevance value recovered from raw per-head attributions.

Source: src/invera_sdk/completions.py:51

ChatMessage​

@dataclass(frozen=True)
class ChatMessage

One chat message rendered through a profile-specific tokenizer template.

Source: src/invera_sdk/completions.py:39

ChatMessageRole​

ChatMessageRole = Literal['assistant', 'system', 'tool-description', 'tool', 'user']

Roles accepted in a chat request. tool-description messages are display-only: the default tokenizer never renders them into the prompt, so supply tool definitions through tools.

Source: src/invera_sdk/completions.py:35

ChatTool​

ChatTool = dict[str, Any]

An OpenAI-compatible function tool definition accepted by the chat template.

Source: src/invera_sdk/completions.py:47

ClaimEvidence​

@dataclass(frozen=True)
class ClaimEvidence

Evidence and source grounding for one stable assistant claim.

Source: src/invera_sdk/evidence.py:81

ClaimVerificationInput​

@dataclass(frozen=True)
class ClaimVerificationInput

Input passed to a host-provided semantic claim verifier.

Source: src/invera_sdk/evidence.py:68

ClaimVerifier​

ClaimVerifier = Callable[[ClaimVerificationInput], SemanticVerification]

A host extension that can turn attribution grounding into semantic support.

Source: src/invera_sdk/evidence.py:77

CompletionRequest​

@dataclass(frozen=True)
class CompletionRequest

Options for one greedy attributed completion.

Source: src/invera_sdk/completions.py:76

CompletionStreamChunk​

@dataclass(frozen=True)
class CompletionStreamChunk

A compatibility chunk matching the former invera Engine chat-completions stream.

Source: src/invera_sdk/completions.py:62

EngineCompatibilityError​

class EngineCompatibilityError(ValueError)

Raised when an Engine cannot safely serve this SDK.

Source: src/invera_sdk/__init__.py:72

EvidenceOptions​

@dataclass(frozen=True)
class EvidenceOptions

Options controlling claim boundaries, verification, context, and local tool policy.

context decides which messages count as grounding for claims. "auto" (the default) ignores attribution to the question, the last user message, when it is shorter than the rest of the system, user, and tool input; a longer last user message carries the retrieved text itself and still counts. "conversation" counts every message, as SDK 0.7.2 and earlier did.

Source: src/invera_sdk/evidence.py:158

EvidenceSource​

@dataclass(frozen=True)
class EvidenceSource

A context source selected from word attribution for an evidence object.

Source: src/invera_sdk/evidence.py:46

EvidenceStreamEvent​

@dataclass(frozen=True)
class EvidenceStreamEvent

One event from the semantic completion stream with evidence.

An output_kind event is emitted once, as soon as the kind of the FIRST output is observed (definitive is true because the kind was directly observed, not heuristically guessed). A response may still continue with the other kind afterwards; the final ResponseEvidence.output_kind then reports "mixed".

Source: src/invera_sdk/evidence.py:186

EvidenceVerdict​

EvidenceVerdict = Literal['supported', 'contradicted', 'grounded', 'partially_grounded', 'ungrounded', 'unknown']

How strongly the available context grounds or verifies one claim.

Source: src/invera_sdk/evidence.py:27

InveraClient​

class InveraClient

Lazy high-level invera Engine client compatible with the old attributed chat stream.

Source: src/invera_sdk/completions.py:193

InveraClientOptions​

@dataclass(frozen=True)
class InveraClientOptions

Construction options for the high-level attributed-completion client.

Source: src/invera_sdk/completions.py:97

InveraHttpError​

class InveraHttpError(RuntimeError)

Raised when an Engine HTTP request fails.

Source: src/invera_sdk/__init__.py:76

ModelInfo​

@dataclass(frozen=True)
class ModelInfo

Stable model details intended for application display and diagnostics.

Source: src/invera_sdk/completions.py:87

OutputKind​

OutputKind = Literal['text', 'tool_call', 'mixed']

The semantic kind observed in an assistant response.

Source: src/invera_sdk/evidence.py:24

ParameterEvidence​

@dataclass(frozen=True)
class ParameterEvidence

Evidence connecting one generated tool argument to conversation context.

Source: src/invera_sdk/evidence.py:110

ResponseEvidence​

@dataclass(frozen=True)
class ResponseEvidence

Final evidence report emitted after every claim and tool call is assessed.

Source: src/invera_sdk/evidence.py:175

SchemaIssue​

@dataclass(frozen=True)
class SchemaIssue

A machine-readable JSON-schema problem found in a generated tool call.

Source: src/invera_sdk/evidence.py:120

SemanticVerification​

@dataclass(frozen=True)
class SemanticVerification

Optional semantic judgement supplied by a host verifier.

Source: src/invera_sdk/evidence.py:59

TextSpan​

@dataclass(frozen=True)
class TextSpan

Character offsets into the visible assistant text.

Source: src/invera_sdk/evidence.py:38

ToolCallEvidence​

@dataclass(frozen=True)
class ToolCallEvidence

A completed tool call assessed before an agent harness executes it.

Source: src/invera_sdk/evidence.py:137

ToolPolicy​

@dataclass(frozen=True)
class ToolPolicy

Local action-risk metadata that is never rendered into the model prompt.

Source: src/invera_sdk/evidence.py:100