#Agents and runtime
| Chinese term | English | Definition | Usage in iGraph |
|---|---|---|---|
| 智能体 | Agent | A software actor that combines model reasoning, task state, and tool calls to pursue a goal. | iGraph is a graph data agent: it selects evidence-producing operations from user questions, reads graph queries, algorithm and learning results, and organizes further analysis and interpretation. Graph Engine performs the computation. |
| 技能 | Skill | Reusable instructions, workflows, and supporting resources organized around a task family for an agent to apply when needed. | Skills guide analysis and construction workflows. The construction agent uses domain experience to choose transformations, validates candidates on fixed queries, and reuses learned experience in later strategy selection. A skill is not simply an alias for an algorithm function. |
| 工具调用 | Tool call | A request for a defined operation using its declared name and parameter structure, followed by a structured result. | For example, shortest_path accepts source and target. The agent calls Graph Engine through MCP, using success, result fields, and errors to determine the next step. |
| 会话 | Session | An isolated scope for an interaction and its associated state. | Within the default workspace, iGraph uses session_id to scope the active graph, uploaded files, models, and jobs. Switching sessions is different from switching graph_version within a session; models must also resolve within their session scope. |
| 后台作业 | Asynchronous job | An execution unit that continues after submission and exposes status, progress, and a final result through an identifier. | Long tasks such as training may return job_id. get_job_status queries the job and exposes the completed result under job.result. Requesting cancellation and confirming that computation has stopped are distinct stages. |
| 服务器发送事件 | Server-sent events (SSE) | A streaming mechanism in which a server sends events to a client over a persistent HTTP response. | iGraph streams runtime messages and progress to the interface. Event arrival, connection closure, and a terminal job state are distinct facts; the interface should use explicit state fields to determine completion. |
| 用户记忆 | User memory | Preferences and context saved for a user and reused in later interactions. | Agentic User Memory uses update_user_memory; the memory panel reads the user-memory store. Verify the stored list and recall in a new session rather than relying only on an assistant acknowledgement. |
| 可复用经验 | Learned knowledge | Knowledge or experience distilled from tasks for retrieval on related future questions. | save_learning writes searchable Learned Knowledge, separate from update_user_memory and construction domain experience. Successful knowledge retrieval does not mean the entry appears in the user-memory panel. |
#Graph structure and algorithms
| Chinese term | English | Definition | Usage in iGraph |
|---|---|---|---|
| 多重边 | Parallel edges / multiedges | Multiple distinct relationships between the same endpoint pair, potentially with different identities, types, or attributes. | GraphData preserves logical edge identity. Some algorithms require a temporary projection, such as the minimum weight per endpoint pair for shortest paths. This calculation view does not delete or permanently merge source relationships. |
| 最短路径 | Shortest path | A path minimizing total cost from source to target under the chosen direction and cost model; unweighted cost is usually the edge count. | shortest_path interprets weights as distances or costs when use_weight=true, and minimizes hops when false. If weights represent similarity, first establish whether they are suitable as direct path distances. |
| 社区发现 | Community detection | Identifying node groups that are relatively densely connected internally and more sparsely connected to other groups. | community_detection supports methods such as Louvain and returns community sizes, node previews, and modularity. Communities are structural groups rather than automatic business labels, and results depend on direction, weights, and algorithm choices. |
| 介数中心性 | Betweenness centrality | The share of shortest paths between other node pairs that pass through a node, describing its intermediary role in shortest-path structure. | Select centrality_type=betweenness in node_centrality to identify potential bridging nodes. High betweenness is a structural property under a particular graph and path definition, not direct evidence of real-world business influence. |
| 接近中心性 | Closeness centrality | A measure of proximity based on shortest-path distances to reachable nodes, commonly related to the reciprocal of mean distance. | Select centrality_type=closeness in node_centrality. Interpret values using connectivity, direction, and distance semantics; the metric is not just degree, and unreachable nodes must not be treated as zero-distance nodes. |
| 全局聚类系数（传递性） | Global clustering coefficient (transitivity) | Under undirected simple-graph semantics, the proportion of connected two-step triples closed into triangles: three times the triangle count divided by the number of connected triples. | global_metrics with metric=transitivity computes this value on an undirected calculation view. It differs from average_clustering, which averages local clustering coefficients across nodes. |
#Construction, validation, and versions
| Chinese term | English | Definition | Usage in iGraph |
|---|---|---|---|
| 图版本 | Graph version (graph_version) | An identifier for a graph state that can be saved, loaded, and analyzed within a session. | get_graph_info returns the active graph_version, and load_graph_version selects a saved one. Graph ML uses it to check input consistency. Dataset construction revision and execution generation are separate identifiers. |
| 构图修订 | Construction revision | A revision number for a constructed dataset document as it is edited and committed. | Construction edits carry base_revision, checked against the current revision to prevent stale edits overwriting newer content. load_constructed_graph loads the constructed result into the session graph; revision must not be substituted for graph_version. |
| 执行代次 | Execution generation | A control identifier distinguishing the current valid execution of a run from stale executions. | Run, job, and worker receipts carry generation. Execution control validates generations when continuing work and publishing state to reject stale results. It describes execution validity rather than changes to graph content. |
| SQL 参考执行器 | SQL reference executor (oracle) | An executor that evaluates queries against independent reference data to produce comparison answers. | Construction validation compiles fixed query plans to SQL against source tables, then compares those answers with GraphExecutor reading only the candidate graph. oracle denotes the validation role, not the Oracle database or model-authored answers. |
| 源数据完整性 | Source-data integrity | The preservation of source-record identity, field values, and declared relationship semantics after a graph schema transformation. | source_integrity_errors checks every source record and declared foreign key, including data untouched by the fixed queries. Passing all queries and passing integrity checks are separate requirements. |
| 固定查询工作负载 | Frozen query workload | A set of queries and execution plans fixed before candidate comparison and kept unchanged throughout that comparison. | The construction agent submits queries with set_workload, then compares semantic correctness and execution cost on that same workload. Freezing prevents score improvement by changing the questions; pass rate applies to this set, not arbitrary questions. |
#Graph learning
| Chinese term | English | Definition | Usage in iGraph |
|---|---|---|---|
| 消息传递 | Message passing | A graph-neural-network computation that collects information along edges, aggregates it, and transforms node representations. | Registered GNN backbones use x and edge_index for node representations; APPNP propagates after prediction. Current model calls do not pass edge weights or xe into the forward computation. Propagating along edges does not mean using all edge attributes. |
| 图级读出 | Graph-level readout | An operation that aggregates representations of nodes or edges within a graph into a fixed-dimensional graph representation. | Graph classification uses global_mean_pool to average each graph's node embeddings into a [B, D] batch representation before a graph-level linear head. The current interface does not expose a switch to sum or maximum readout. |
| 嵌入向量 | Embedding vector | A fixed-dimensional numerical representation of an object such as a node or graph for prediction and similarity computation. | Pretraining saves node embedding.npy and corresponding names. data.feature_source=embedding reuses these vectors by name. Reusing embeddings does not restore an upstream training checkpoint; unmatched node names currently receive zero vectors. |
| 节点分类 | Node classification | A supervised task that assigns a class label to each target node. | task=node_classification runs a GNN and classification head on the active graph, applying cross-entropy to labeled training nodes. Two-dimensional node labels are reduced with argmax, so this remains single-label rather than multi-label node classification. |
| 链接预测 | Link prediction | A task that scores node pairs for their likelihood of being connected. | task=link_prediction uses an embedding dot-product decoder and returns scores and sigmoid probabilities. The decoder is symmetric, and validation edges currently remain in message passing; both conditions matter when interpreting metrics. |
| 图分类 | Graph classification | A task that predicts one or more labels for an entire graph using a graph-level representation. | task=graph_classification uses a separate multi-graph dataset: one binary task for HIV and 128 binary tasks for PCBA. Node message passing is followed by graph-level readout. A graph label and labels of individual nodes are different supervision targets. |
Terminology references: the official documents below provide Chinese and English usage. iGraph's data structures, algorithm options, and runtime boundaries are defined in this project's corresponding technical chapters.