Context
Budgeting, parallel retrieval, ranking, packing, formatting, and safe degradation for LLM context.
Context assembly converts directives and retrieved memories into a bounded prompt package. The pipeline exists locally. Caller authentication and the accepted deadline boundary are not complete, so the service is restricted to loopback or a trusted network boundary.
Provisional Ledger class: provisional
Verified 2026-10-21Caller authentication and deadline acceptance remain open
Pages
- Ranking and packing — retrieval deadlines, budget formula, scoring, DP/greedy packing, and formatting safety.
Pipeline
Rendering diagram
- Budget — model capability catalog and a character/rune estimate. This is not exact tokenizer accounting.
- Retrieve — directive Redis, hot memory, and cold semantic memory run in parallel with independent failure handling.
- Score — interim similarity and confidence scoring; category half-lives are applied by the memory substrate.
- Pack — bounded dynamic-programming knapsack with a greedy fallback.
- Format — deterministic order and nonce-delimited, escaped memory serialization.
retrieval_wall_ms = min(timeout_ms, deadline_ms). The defaults are 45 ms for retrieval and 40 ms for the gRPC deadline.
Security limitation
The gRPC server has no auth interceptor. On a non-loopback bind it trusts request
org_id and agent_id. Do not expose it to an untrusted network. A network address is not an authentication control.Keep IBEX_CONTEXT_GRPC_ADDR=127.0.0.1:<port> unless a documented trusted-sidecar boundary exists. Do not claim caller authentication until the deferred interceptor lands.
Failure behavior
| Failure | Result |
|---|---|
| Directive lookup fails | Continue with memory branches |
| Hot or cold memory fails | Continue with available branches |
| Retrieval exceeds wall budget | Return partial or fallback context |
| Packing cannot fit | Greedy fallback, still bounded |
| Context service unavailable | Proxy proceeds with directive-only behavior |
Context degradation is safe only if the caller cannot force cross-tenant retrieval. That is exactly why the auth limitation is called out above.
Explicitly not implemented
SearchMemoriesandRecordMemoryFeedbackreturnUNIMPLEMENTED.- Caller identity propagation from the proxy is not authenticated at the context boundary.
- The character estimate is not a provider-exact tokenizer.
- There is no accepted hosted latency or quality benchmark for this service.