Module dsv41_encoding 

DeepSeek-V4.1 prompt encoding and completion parsing.

The V4.1 checkpoint intentionally ships without a Jinja chat template. This module is a Rust port of the official encoding/encoding.py at the pinned model revision. It keeps the protocol in one place for the CLI, OpenAI server, and any embedding application: spaced DSML tags, numeric reasoning effort, mid conversation system turns, tool result merging, and image ordering all follow the reference implementation.

Structs

EncodeOptions

EncodedPrompt

Enums

EncodingError

ReasoningEffort

ThinkingMode

Constants

ASSISTANT_SP_TOKEN

BOS_TOKEN

DSML_TOKEN

EOS_TOKEN

IMAGE_PLACEHOLDER

LATEST_REMINDER_SP_TOKEN

SYSTEM_SP_TOKEN

TASK_TOKENS

Internal quick instruction tasks supported by the official formatter.

THINKING_END_TOKEN

THINKING_START_TOKEN

USER_SP_TOKEN

Functions

encode_case

encode_messages

Encode current messages. images contains only records belonging to the current turn, matching the official helper’s context-cache contract.

encode_messages_simple

merge_tool_messages

Merge standalone OpenAI tool messages into a preceding user content list.

parse_message_from_completion_text

Parse a generated V4.1 completion into an OpenAI style assistant object.

parse_tagged_text

Convert compact <image>path</image> input to OpenAI style content blocks.

parse_tool_calls

process_image_messages

Normalize image blocks and return image records in prompt order.

render_message

Render one message exactly as the official Python implementation does.

sort_tool_results_by_call_order

Stable tool result order follows the preceding assistant’s tool call order.

to_json