The words used in the product, and the boundary between saved configuration and a finished run.
Core concepts
| Term | Meaning |
|---|---|
| Workspace | The boundary for members, role, retaining policy, AI model choices, and embedding default |
| Catalog collection | A named group of documents and their extracted output in a stored workspace |
| Schema | The fields extraction should fill. A proposal is not an approved schema |
| Extraction run | A saved execution against an approved configuration. Status moves through processing to a terminal state |
| Smart Bite | A focused main passage with its source. Multi-size chunking can add smaller sentence groups and larger multi-chunk windows (500 / 1,500 / 4,500 by default) that project back onto main Smart Bites |
| Hybrid retrieval v2 | Collection search that fuses PostgreSQL-native pgvector dense search, Hebrew/English BM25 keyword search, section lookup, multi-pass queries, and optional extra chunk sizes |
| Answer and brief | A grounded reply written in the question’s language (question-language/v1) and structured as a cited brief (answer-brief/v1: Answer, Prevailing position, Direct support, Supporting context, The other side, Gaps), or an explicit status such as insufficient evidence |
| Studied field answer | A deterministic list, count, ranking, or aggregate across a collection’s normalized extracted field values (field_answers_v1), where every value carries a verified passage span citation and field coverage is stated |
| Table facts | Reconstructed logical tables (tables-v3), cell-level numeric and date facts ([TF1]), and exact arithmetic reconciliation findings ([RF1]) checked by the built-in finance rule pack |
| Bitemporal graph | Facts and entities tracked across two time axes in Lineage and Context Studio: True at (valid time in the world) and As known at (system time when Bundata recorded the document or fact) |
| Context Engine | The workspace and collection control plane (/platform/context-engine) that tracks coverage across 9 derived knowledge layers, runs dry-run/build/resume jobs, gates advanced answer paths with Eval Studio evidence, and manages ranking profiles |
| Context Studio | The workspace hub (/platform/context-graph) containing Explore (chat beside a live canvas), Evidence, Quality review (bitemporal correction queues), Library (approved business terms, verified answers, instructions, and SQL query templates), Context sets, Data (PostgreSQL/MySQL/dbt metadata), and Enrich (Gaps and enrichment proposals) |
| Context set | A versioned, steward-governed bundle of approved library entries, rule templates, and scope (collections or documents) that agents start from |
| Agent credential and memory | A scoped bdk_… token for /api/agent/v1 and MCP. When memory:read/memory:write and the owner setting are enabled, agents can keep PII-redacted, source-checked notes for their creator (evidence: false, trust: "untrusted") |
| Connection | Saved source (S3, GCS, Azure Blob, Google Drive, PostgreSQL, MySQL) or destination settings. Registration is not a successful read, and a read is not a write |
Uploaded, extracted, indexed, and knowledge-built are distinct states. A confirmed upload can still be processing. An extracted Smart Bite can still be waiting for vector indexing or a Context Engine layer build.
See the Glossary for the Japanese labels used in the application.