Glossary
A shared vocabulary for collections, chunks, models, and tools.
Use these terms to distinguish the archive's data model from generation, speech, and network configuration. The same word can have a different meaning in a provider's export or a client application; the definitions here describe ThreadShelf Enhanced.
Archive vocabulary
| Term | Meaning |
|---|---|
| Export | A file or archive produced by an AI service or a supported converter; the input to import |
| Collection | A named namespace for conversations and their searchable content |
| Conversation / thread | A complete discussion identified by collection, source file, and conversation key |
| Turn | A normalized user, thinking, or assistant text entry, with available metadata |
| Chunk / passage | A searchable portion of a turn; long turns can create several chunks |
| Source file | The stored identity of the input containing a conversation, not necessarily a file the reader must reopen for every view |
| Conversation key | The identity that distinguishes conversations within a source; clients should retain the returned value |
| Raw data | The retained original conversation structure alongside normalized turns; not a guarantee of byte-identical input restoration |
| Continuation | New ThreadShelf-authored turns appended to an imported conversation |
| Protected collection | An application-reserved collection with restrictions on bulk clearing or deletion |
See Collections and insights for counts and lifecycle behavior and Compatibility for identity contracts.
Search vocabulary
| Term | Meaning |
|---|---|
| Embedding | A numerical representation of a passage or query used for semantic retrieval |
| Semantic search | Retrieval by similarity between local query and passage embeddings |
| Exact / keyword search | Literal text matching; Exact is the interface label and keyword is the API/CLI mode name |
| Hybrid search | Independent lexical and semantic retrieval combined into a ranked result list |
| Lexical cache | A disposable local SQLite search cache derived from published conversation content |
| Distance | A semantic retrieval value; a lower cosine distance indicates a closer vector, not a factual accuracy score |
| Origin filter | A filter distinguishing imported archive turns from ThreadShelf-authored turns |
| Reader snapshot | A bounded, expiring result snapshot used by the compatible paginated Reader API |
A search result is a passage, so several results can open the same thread. Search and discovery explains practical mode choice and ranking.
Import and storage vocabulary
Transfer moves browser-selected file bytes to the server. Acceptance is the point at which the server has taken responsibility for a durable import job. These are separate events: do not close the upload tab before acceptance.
Staging is temporary disk-backed preparation of a candidate import. Publication makes a matching archive/index generation active. Here, “generation” means a storage revision, not text produced by an AI model.
Replacement publishes a complete new input selection for a collection after preparation. Additive import can add or refresh sources without treating that selection as the entire collection. Cancellation explicitly asks a job to stop; closing an observer connection does not cancel an already accepted job.
RSS is resident process memory, including memory beyond JavaScript's heap. The memory governor considers process and host/container budgets at work boundaries. Increasing the JavaScript heap alone does not create more physical RAM or guarantee that native allocations fit.
See Import jobs and Architecture.
AI and speech vocabulary
| Term | Meaning |
|---|---|
| Provider | A generation or speech backend; it may run locally, on a self-hosted service, or through an external API |
| Runtime | The software that executes a local model, such as a supported llama.cpp runtime |
| GGUF | A model-file format used by supported local generation workflows |
| Context window | The input-and-output token capacity available to a model request; it is distinct from total archive size |
| Master prompt | A saved system prompt selected for generation; it is not rewritten into every stored thread |
| Private chat | A tab-scoped conversation kept in browser memory instead of the ordinary saved-chat workflow |
| Tool | A bounded capability, such as archive search or an approved note edit, available to a model/client within configured scope |
| Vault | A connected live Markdown folder with its own identity and read/write policy |
| TTS | Text-to-speech: converting selected text into audio |
| Pinned recording | Saved audio protected from ordinary cache eviction and replacement policies |
| Streaming | Delivering text or audio incrementally; it does not by itself mean that a complete result was saved |
Private mode changes local persistence. It does not make a remote provider local, bypass access checks, or revoke a separately approved note write. See Privacy and security.
Network vocabulary
Loopback addresses, such as 127.0.0.1, refer to the machine making the request. Bind address determines where the server listens. Allowed host validates the requested hostname or address, while browser origin includes scheme, host, and port. None of these alone identifies a person.
LAN feature permission enables a corresponding group of AI or speech controls at allowed server addresses. It is not an account system. A local model API key protects the separate inference API according to its configured policy; it is not a universal login for the archive web interface.
MCP stdio connects a client and a locally launched tool process over standard input/output. It is different from the HTTP archive API and the OpenAI-compatible /v1 inference routes.
For setup and examples, see Network access, Archive HTTP API, Local model API, and MCP.
Source references
Behavior is documented from the Enhanced source. For changes after this handbook was built, check the project changelog.