Skip to content

Commit 8e3778f

Browse files
committed
Fix inconsistencies leftover from new section additions.
Signed-off-by: usize <mofoster@redhat.com>
1 parent 371dd51 commit 8e3778f

1 file changed

Lines changed: 25 additions & 13 deletions

File tree

docs/plans/ai-gateway.md

Lines changed: 25 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -68,9 +68,9 @@ networking in Gateway API. Two proposals are directly relevant:
6868
- [Proposal 7: Payload Processing] — introduces
6969
`PayloadProcessingPipeline` for ordered, sequential body/header
7070
processors (prompt validation, PII redaction, semantic routing) as
71-
HTTPRoute filters. This is the emerging standard for guardrails and
72-
is the intended generation target for per-provider processing
73-
pipelines (see [Future: guardrails](#future-guardrails)).
71+
HTTPRoute filters. This is the emerging standard for guardrails.
72+
How we integrate it with our abstraction is discussed in
73+
[Future: PayloadProcessingPipeline](#future-payloadprocessingpipeline).
7474

7575
Our design generates the Envoy AI Gateway CRDs that exist today, but is
7676
structured so that when the WG proposals mature into accepted APIs, we
@@ -737,10 +737,9 @@ presents its SPIFFE certificate (CSI driver, spiffe-helper sidecar,
737737
AuthBridge integration, or native go-spiffe) is the client-side
738738
concern and is out of scope for this proposal.
739739

740-
AuthBridge (agent-to-agent OAuth/OIDC) and the MCP Gateway (MCP
741-
protocol routing) are orthogonal to AI Gateway (model routing and
742-
inference access control). They target different traffic flows and
743-
do not conflict.
740+
AuthBridge (agent-to-agent OAuth/OIDC) is orthogonal to AI Gateway
741+
(model routing and inference access control). They target different
742+
traffic flows and do not conflict.
744743

745744
Client-side configuration for connecting workloads to AI Gateways
746745
with SPIFFE identity may be addressed in a separate proposal.
@@ -849,12 +848,26 @@ config:
849848
hooks:
850849
xdsTranslator:
851850
post: [Translation, Cluster, Route]
851+
translation:
852+
cluster:
853+
includeAll: true
854+
listener:
855+
includeAll: true
856+
route:
857+
includeAll: true
858+
secret:
859+
includeAll: true
852860
service:
853861
fqdn:
854862
hostname: ai-gateway-controller.<ns>.svc.cluster.local
855863
port: 1063
856864
```
857865

866+
The `translation` block is required. Without it, Envoy Gateway
867+
defaults to excluding listeners and routes from the
868+
`PostTranslateModify` hook, and the AI Gateway extension server
869+
cannot inject the ext_proc filter into the listener filter chain.
870+
858871
## Implementation plan
859872

860873
### Phase 1: AIRoutingPolicy
@@ -894,13 +907,12 @@ enable provider reuse across gateways and finer-grained RBAC. The
894907
internal `RoutingIntent` model should be structured so that this
895908
extraction is a mechanical refactor.
896909

897-
### Future: guardrails
910+
### Future: payload processing
898911

899-
Per-provider guardrails via the `processing` field (see
900-
[Future: guardrails](#future-guardrails) above). Depends on the
901-
WG AI Gateway [Proposal 7: Payload Processing] maturing. Gateway-wide
902-
guardrails would be a separate policy CRD and a separate design
903-
proposal.
912+
Per-provider and per-model guardrails via
913+
[PayloadProcessingPipeline](#future-payloadprocessingpipeline).
914+
Depends on the WG AI Gateway [Proposal 7: Payload Processing]
915+
maturing.
904916

905917
## Compatibility notes
906918

0 commit comments

Comments
 (0)