Functions
aggregate_completion
def aggregate_completion(raw: Iterator[tuple[int, dict[int, float]]], tokenizer: TokenizerAdapter, request: CompletionRequest, prompt_ids: Sequence[int], eos_token_ids: Sequence[int] | None=None) -> Iterator[CompletionStreamChunk]
Map raw source positions into the former word-level completion stream.
Source: src/invera_sdk/completions.py:432
create_client_with_tokenizer
def create_client_with_tokenizer(options: InveraClientOptions, tokenizer: TokenizerAdapter) -> InveraClient
Create a high-level client with an advanced host-managed tokenizer.
Source: src/invera_sdk/completions.py:390
create_hugging_face_tokenizer
def create_hugging_face_tokenizer(identity: TokenizerIdentity) -> TokenizerAdapter
Load the exact Hugging Face tokenizer revision and accept it only if its fingerprint matches.
Source: src/invera_sdk/completions.py:178
decode_raw_attribution_chunk
def decode_raw_attribution_chunk(payload: Any) -> tuple[int, dict[int, float]]
Decode one raw NDJSON object into a token id and summed source-position scores.
Source: src/invera_sdk/completions.py:399
parse_qwen_tool_calls
def parse_qwen_tool_calls(chunks: Iterator[CompletionStreamChunk], tool_parser: str) -> Iterator[CompletionStreamChunk]
Split Qwen <tool_call> markers into typed compatibility chunks.
tool_parser is one of SUPPORTED_TOOL_PARSERS: the Engine's
sdk_hints.recommended_tool_parser for the model that produced chunks.
Source: src/invera_sdk/completions.py:468
stream_evidence
def stream_evidence(chunks: Iterator[CompletionStreamChunk], request: CompletionRequest, options: EvidenceOptions | None=None, *, tool_parser: str | None) -> Iterator[EvidenceStreamEvent]
Add claim and tool-call evidence to an attributed completion stream.
tool_parser names the format of the tool calls in chunks: the Engine's
sdk_hints.recommended_tool_parser for the model that produced them, one of
SUPPORTED_TOOL_PARSERS. None (the model recommends no parser) accepts only streams
without tool calls.
Source: src/invera_sdk/evidence.py:238