QilbeeDB documentation
Build agent applications with durable memory, scoped retrieval and traceable learning records. QilbeeDB stores evidence and retrieves it under explicit identity and revision contracts. Your application chooses the language model, generates embeddings externally and decides how retrieved evidence is used.
Choose how to use QilbeeDB#
| QilbeeDB platform | Self-hosted installation | |
|---|---|---|
| Connect | Use your company’s API credentials with https://api.qilbeedb.io |
Use the base URL and credentials from your installation administrator |
| Manage your company | Use the administration console and authorized administration APIs | Use the administration capabilities available in your installation |
| Operate the service | Infrastructure and server updates are managed by the platform | Your organization manages hosting, storage, TLS, backups and server updates |
| Build an application | Follow the quickstart | Install QilbeeDB, then follow the same quickstart |
Both options use the same memory and authorization contracts for the server version they run. Your organization controls its application credentials and access policies in either option. Model selection, external embeddings and agent execution remain responsibilities of your application.
Start building#
- What’s new in 0.14.0: review available capabilities, compatibility and upgrade requirements.
- Quickstart: create a memory and retrieve it with a scoped credential.
- Authentication and scopes: understand tenants, resource grants and private subjects.
- Automatic agent registration, 0.12.0: register external IDs on successful authorized requests and inspect the company directory.
- Agent display names (0.14.0): recognize external agent IDs and preserve revisioned name changes.
- Company integrations, 0.12.0: authorize external project and agent IDs through a versioned policy and change access without replacing keys.
- Company memory administration, 0.12.0: discover retained workspaces and inspect company records independently of credential grants.
- Memory API: create, update and delete records with durable receipts and revision checks.
- Local Docker deployment: run the platform and locate its API reference.
Choose a retrieval mode#
| Mode | Input | Ranking signal | Use when |
|---|---|---|---|
| Lexical | Text | BM25 | Exact words and identifiers matter, or no embedding is available |
| Semantic | External vector and model identity | Cosine | Meaning should be matched through your selected embedding model |
| Hybrid, experimental | Text, external vector and ranking version | Weighted reciprocal rank fusion | You want to evaluate complementary lexical and semantic candidates |
| Graph-assisted, experimental | Lexical, semantic or hybrid seeds and a graph profile | Base ranks and the strongest eligible typed path | You want to evaluate explicit relations as an additional retrieval signal |
Each mode isolates the authorized tenant, project, agent, mission and private subject before candidate selection. Responses distinguish ranking scores from probabilities and disclose bounded scan coverage. The hybrid endpoint uses an immutable server-owned profile; it does not change the cosine score returned by the semantic endpoint.
Evaluate retrieval on frozen, judged queries before adopting a ranking version. Report retrieval quality separately from end-to-end agent task outcomes.
Build learning workflows#
Experience receipts, introduced in the 0.7.0 contract, preserve execution intent, authenticated observations and unknown resource consumption. They do not automatically qualify a procedure.
Procedural learning records proposals, evaluation evidence and publication decisions. Learned tools add immutable artifacts, executor profiles and a durable development ledger. Workers execute tool development outside the database process. These contracts make outcomes observable; they do not guarantee that every learned procedure improves an agent or that a generated program is safe merely because it was recorded.
The 0.14.0 company learning catalog adds company-wide discovery and inspection of retained experiences, procedures, strategies and tools, without requiring a memory record or an active writer key.
Use the exact deployed contract#
The server publishes its OpenAPI document at /openapi.json and an interactive
reference at /docs. On the local Docker host, open
http://localhost:7474/docs. Check /health for the
running version. Release-stage metadata in these Markdown sources distinguishes
preview documentation from a validated deployment.
The platform rejects unknown request fields. Vector attachment, cosine, hybrid and graph-assisted search accept JSON bodies up to 2 MiB. Other memory routes retain the 65536-byte limit; login and login-account JSON requests use 8192 bytes. External vector spaces support 1–32768 dimensions, including 3072, subject to the operator capacity configuration. Start with the HTTP contract and errors when integrating a new client. The current default router covers the documented platform endpoints; legacy graph and memory routes are a separate compatibility surface.
Evaluate experience evidence explains how to freeze observations, preserve incomplete outcomes and compare agent tasks.
Use the memory change feed to reconcile scoped caches and resume from a durable cursor after disconnection.
Use memory review to record decisions and exclude rejected revisions while preserving historical evidence.
Use derived memories to keep conclusions dependent on exact source revisions and transitive source eligibility, with bounded diagnostics.
Use consumer checkpoints to persist each subject’s feed progress and reconnect safely after a process restart.
For embedded Rust graphs, see the 0.12.0 durable graph lifecycle contract. The graph memory evidence map explains which research informs the next memory integration and what remains unimplemented.
Use the 0.12.0 memory evidence graph API to read revision-bound ancestry in a single authorized snapshot, including native company administration and explicit traversal coverage.
The 0.13.0 typed memory relation API stores semantic, entity, temporal, causal, support and contradiction assertions with exact endpoint revisions, declared provenance and separate review authority. Applications can consume the relation feed and explicitly select graph-assisted retrieval. Existing search routes do not automatically traverse these assertions.
The typed graph API traverses those assertions in either direction, returning current eligible endpoint records and explicit work coverage. Company administrators can also inspect retained assertion state and history directly.
Use the relation change feed to invalidate graph caches and retain consumer progress across interruptions. Its history-bound cursors and explicit reconciliation complement the memory feed; current eligibility checks remain necessary before cached context is reused.
The graph-assisted retrieval API selects lexical, semantic or hybrid anchors and ranks memories through exact typed paths in one snapshot. Four immutable experimental profiles expose separate base and graph contributions, coverage and provenance. Existing search scores and defaults remain unchanged; the reserved comparison found uncertain mean improvement and material regressions. Agent-task benefits remain unmeasured.
The 0.14.0 learning evidence history lets administrators inspect original evaluation submissions and retained events from a selected learning resource, with explicit historical states and recovery.
The 0.14.0 company consumer directory lets administrators discover retained memory and graph consumers across company projects and private owners, then inspect progress without a writer's key. Legacy sequence estimates remain explicitly unverified, and saved checkpoints do not prove worker liveness. The console supports selection, continuation, safe retry, and keyboard return without requiring manual consumer identifiers.
The 0.14.0 external graph consolidation contract coordinates durable external extraction, exact source revisions and fenced atomic relation publication. Its worker and administrative workflows require qualification before release; inferred relations do not establish an agent-quality gain.