Glossary | LurusTech Docs

Glossary

Key terms from the Lurus product docs — 71 entries, grouped by topic. Use Ctrl + F to jump quickly.

General

Concepts shared across products, API, and infrastructure

鹿贝: Lurus's unified billing unit (Lubei). All product usage is charged in Lubei; 1 Lubei equals one US cent.
Agent: An LLM-driven autonomous program: perceive input → reason → call tools → produce results.
API Key: A secret credential for accessing the API, much like a password. Each Key can have its own quota and permissions. Read details →
ArgoCD: A GitOps continuous-delivery tool for Kubernetes that watches a Git repository for changes and syncs them to the cluster automatically.
Base URL: The root address of the API. The Lurus Base URL is https://hub.lurus.cn/v1.
Bearer Token: An HTTP authentication scheme that carries the credential in the request header as "Bearer sk-xxx".
Embedding: A vector that converts text into a high-dimensional numeric representation, so semantically similar text lies close together in vector space. MemX generates these with a local ONNX model.
Endpoint: An API endpoint — a callable URL path such as /v1/chat/completions.
FIFO: First In, First Out. Kova's task scheduling uses a FIFO strategy by default.
GitOps: Using a Git repository as the source of truth for deployments, with automation tools such as ArgoCD syncing the desired state to the cluster.
JWT: JSON Web Token — a self-contained, signed token that lets the server verify identity without a database lookup.
Ontology: A tree-structured description of domain knowledge (user stories, architecture, tech stack, conventions). Forge's core data model.
OpenAPI: An open specification for describing REST APIs (formerly Swagger). Lurus data structures are auto-generated from the gateway's OpenAPI spec. Read details →
OpenTelemetry: An open standard for observability that unifies the collection and transport of the three signals — traces, metrics, and logs.
PAT: Personal Access Token — a long-lived API access token generated by the user to replace passwords for automation.
PII: Personally identifiable information, such as API Keys, passwords, and private keys. MemX ships with 12 PII filter rules that cannot be bypassed.
Rate Limit: A cap on the number of requests allowed per unit of time, used to prevent abuse and overload.
Schema: A definition of a data structure that specifies the name, type, and required status of each field in a request or response. Read details →
SDK: A software development kit providing ready-made functions and classes to simplify API calls. The Lurus gateway uses the common Chat Completions interface format.
SM4-GCM: The Chinese-national SM4 block cipher in GCM authenticated-encryption mode — the go-to symmetric encryption for SM-cryptography compliance scenarios.
Streaming: A streaming response where the AI model returns results token by token rather than waiting for full generation, reducing time to first token.
System Prompt: A system prompt that defines the AI's role, capability boundaries, and behavioral rules. It goes first in the message array.
TCO: Total Cost of Ownership — the full-lifecycle cost covering licensing, deployment, operations, and migration in a procurement decision.
Token: The smallest unit an LLM uses to process text. Roughly 1.5 characters per token for Chinese and 4 characters per token for English.
Tool Call: When an AI model decides during reasoning to call an external tool (such as search or a database query) and returns structured call arguments.
Webhook: A server-initiated push of event notifications to a URL you specify, such as payment success or subscription expiry.

Authentication

Bearer Token: An HTTP authentication scheme that carries the credential in the request header as "Bearer sk-xxx".
Casdoor: The open-source identity platform used by Lurus, providing OIDC / OAuth2 / SAML / SCIM / MFA capabilities.
JWT: JSON Web Token — a self-contained, signed token that lets the server verify identity without a database lookup.
MFA: Multi-factor authentication, which requires two or more credentials at once (password + TOTP, password + Passkey, etc.).
OIDC: OpenID Connect, an identity authentication standard built on OAuth 2.0. Lurus implements it with Casdoor.
Passkey: A passwordless authentication method based on the WebAuthn standard, completed via device fingerprint, Face ID, or a security key.
PAT: Personal Access Token — a long-lived API access token generated by the user to replace passwords for automation.
SM4-GCM: The Chinese-national SM4 block cipher in GCM authenticated-encryption mode — the go-to symmetric encryption for SM-cryptography compliance scenarios.
SSO 联邦: Enterprises connect their own IdP (Okta, Azure AD, etc.) to Lurus Auth so employees can sign in to Lurus products with their corporate accounts.

Kova Execution Engine

Agents, WAL, MCP, and scheduling

A2A: The Agent-to-Agent protocol, which lets multiple agents communicate directly, delegate tasks, and exchange information.
Agent: An LLM-driven autonomous program: perceive input → reason → call tools → produce results.
Agent Loop: An agent's decision loop: receive input → LLM reasoning → call tools → get results → keep reasoning, until a final answer is produced.
Checkpoint: A complete in-memory state snapshot written to disk. On crash recovery, the WAL is replayed starting from the most recent Checkpoint.
CheckpointSaver: The component interface in LangGraph responsible for persisting graph state; it can be swapped for implementations provided by Kova or Lumen.
CRC32: A cyclic redundancy check algorithm used to detect errors in data transmission or storage. Every Kova WAL record includes a CRC32 checksum.
Durable Execution: Task state is continuously written to the WAL, so the process automatically resumes from the point of interruption after a crash without re-invoking the LLM.
FIFO: First In, First Out. Kova's task scheduling uses a FIFO strategy by default.
MCP: Model Context Protocol — an open standard that lets AI agents call external tools and data sources in a uniform way. Read details →
MCP Server: A tool service process that implements the MCP protocol, exposing callable tool capabilities (GitHub, databases, etc.) to AI agents. Read details →
Replay: Reproducing an agent's execution from a persisted event sequence — debug locally without spending extra LLM tokens.
Ring Buffer: A fixed-size circular storage structure. When the write pointer reaches the end it wraps around and overwrites old data; the WAL uses a power-of-2-sized ring buffer.
Swarm: A swarm-intelligence pattern in which a coordinator assigns tasks to multiple agents that collaborate autonomously, with no predefined fixed workflow. Read details →
System Prompt: A system prompt that defines the AI's role, capability boundaries, and behavioral rules. It goes first in the message array.
Tool Call: When an AI model decides during reasoning to call an external tool (such as search or a database query) and returns structured call arguments.
WAL: Write-Ahead Log. Every state change is written to the log before it executes, ensuring recovery from the log after a crash. Read details →

MemX Memory

Distillation, deduplication, decay, retrieval

ACE: Adaptive Context Engine. The core intelligence layer of MemX, comprising four modules: knowledge distillation, deduplication, decay, and retrieval. Read details →
Cosine Similarity: A measure of how similar two vectors' directions are, ranging from [-1, 1]. MemX uses it to judge whether knowledge is duplicated (>=0.8 merges automatically).
Curator: MemX's semantic-dedup engine, which uses cosine similarity on every write to automatically merge duplicate knowledge and detect contradictions. Read details →
Decay: Memory decay that mimics the human forgetting curve: rarely retrieved knowledge gradually loses weight, while frequently retrieved knowledge becomes more firmly established. Read details →
Ebbinghaus: The Ebbinghaus forgetting curve, the psychological law that human memory decays exponentially over time. MemX designs its memory half-life around it.
Embedding: A vector that converts text into a high-dimensional numeric representation, so semantically similar text lies close together in vector space. MemX generates these with a local ONNX model.
Knowledge Distillation: Extracting refined, reusable knowledge entries from large volumes of conversation. MemX's Reflector engine does this automatically.
Ontology: A tree-structured description of domain knowledge (user stories, architecture, tech stack, conventions). Forge's core data model.
PII: Personally identifiable information, such as API Keys, passwords, and private keys. MemX ships with 12 PII filter rules that cannot be bypassed.
Reflector: MemX's knowledge-distillation engine, which extracts high-value knowledge from conversations via rule-based pre-screening plus LLM refinement. Read details →
Semantic Dedup: Semantic deduplication, which uses cosine similarity to automatically identify and merge knowledge entries that mean the same thing but are worded differently.
Session: In Forge, the context container for a single product discussion, holding the complete timeline of dialogue, decisions, and agent output.
Vector Search: Finding the most semantically relevant results by computing the distance between the embeddings of the query and the stored text.

Lucrum Quant

Backtesting, risk metrics, trading rules

鹿贝: Lurus's unified billing unit (Lubei). All product usage is charged in Lubei; 1 Lubei equals one US cent.
Adjustment: Adjusting historical stock prices for dividends and stock splits so prices across different periods are comparable. Split into forward and backward adjustment.
Backtest: Simulating strategy trades against historical market data to evaluate how a strategy would have performed in past market conditions.
Calmar: The Calmar ratio: annualized return divided by max drawdown. Above 1.0 means returns can cover the maximum drawdown.
Commission: The fee a broker charges per trade. On A-shares it is typically about 0.03%, with a 5-yuan minimum.
CtaTemplate: vnpy's base strategy template, which defines standard lifecycle hooks such as on_bar, on_tick, and on_trade.
Decimal.js: A JavaScript arbitrary-precision decimal library that Lucrum uses to eliminate floating-point errors in financial calculations.
Max Drawdown: The maximum percentage drop from a historical peak to a subsequent trough. Below 20% is acceptable; below 10% is excellent.
Round Lot: The minimum trading unit on A-shares: 1 lot = 100 shares. Purchases must be in whole-lot multiples.
Sharpe Ratio: Excess return earned per unit of risk taken. Above 1.0 is good; above 2.0 is excellent.
Slippage: The difference between the expected and actual fill price. Slippage is more pronounced for large orders or illiquid instruments.
Sortino: The Sortino ratio, similar to the Sharpe ratio but considering only downside volatility (loss risk), so it better reflects an investor's real experience.
Stamp Duty: A tax levied by the state when selling stock. The current A-share rate is 0.05% of the transaction amount.
Stop Loss: A preset loss threshold that automatically closes a position once reached, to limit losses. A basic risk-control measure.
vnpy: An open-source Python quantitative-trading framework. Lucrum's AI-generated strategy code is built on the vnpy CtaTemplate.
Win Rate: The share of profitable trades among all trades. Above 50% is baseline; above 60% is good.

Switch Management

CLI, Provider, MCP orchestration

MCP: Model Context Protocol — an open standard that lets AI agents call external tools and data sources in a uniform way. Read details →
MCP Server: A tool service process that implements the MCP protocol, exposing callable tool capabilities (GitHub, databases, etc.) to AI agents. Read details →
Provider: A model provider, i.e. an upstream model service (a cloud API or a local inference service). Switch supports configuring multiple Providers and routing by rule.
Proxy: An API endpoint (Chat Completions format) that Switch runs locally, forwarding requests to different Providers according to routing rules.

Lumen Observability

Replay, Checkpoint, cost

Checkpoint: A complete in-memory state snapshot written to disk. On crash recovery, the WAL is replayed starting from the most recent Checkpoint.
CheckpointSaver: The component interface in LangGraph responsible for persisting graph state; it can be swapped for implementations provided by Kova or Lumen.
Dependency Guardian: A planned dependency sentinel in Forge that automatically detects interface breakage across Epics and Stories.
Replay: Reproducing an agent's execution from a persisted event sequence — debug locally without spending extra LLM tokens.
Session: In Forge, the context container for a single product discussion, holding the complete timeline of dialogue, decisions, and agent output.

Missing a term?

If a term is missing or needs more detail, feel free to open an Issue. GitHub Issues ↗