pjandro/OmniRoute
1
0
Fork 0
mirror of https://github.com/diegosouzapw/OmniRoute.git synced 2026-08-25 16:42:16 +03:00
Code Issues Packages Projects Releases Wiki Activity

merge #11383 onto updated tip

Browse Source
This commit is contained in:
Markus Hartung
2026-08-24 09:44:36 -03:00
parent 7715825cb8 38d21afc2d
commit d23bfefec0
12 changed files with 599 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
docs/security/GUARDRAILS.md
View File

@@ -7,7 +7,7 @@ lastUpdated: 2026-08-24
# Guardrails
> **Source of truth:** `src/lib/guardrails/`
> **Last updated:** 2026-08-24 — v3.8.50 (Video Bridge visual dedup hardening)
> **Last updated:** 2026-08-24 — v3.8.50 (Video Bridge visual dedup hardening + focused captions)
Guardrails enforce safety, policy, and content transformations at the boundary
between OmniRoute and upstream providers. Each guardrail can inspect (and
@@ -335,6 +335,19 @@ policies are performed only inside the normalized interval. The resulting
window is included in sampling metadata and in the untrusted description
prefix so downstream models can distinguish a focused excerpt from the full
timeline.
Semantic caption focus is a separate, explicit setting. The default `full`
analysis mode preserves the existing frame prompt and never forwards request
text to the caption model. In `focused` mode, the bridge reads only the latest
non-empty user-authored `text`/`input_text` from the same Chat or Responses
container, normalizes it to NFC, collapses control characters and whitespace,
and limits it to 500 Unicode code points. An empty result falls back to the
exact `full` prompt. A usable hint is serialized as JSON in a dedicated
untrusted-user-context block and may only prioritize observable details; it
cannot override the separate warning against following instructions visible
or audible in the media. Textual focus never infers `start`/`end` or changes
the temporal sampler.
Each frame is limited to 4 MiB, all raw frames together to 23 MiB, and the
serialized broker response to 32 MiB. A private temporary directory is removed
in `finally`. OmniRoute does not bundle FFmpeg and does not accept a custom
@@ -408,12 +421,16 @@ including a fallback model; the bridge reports `mixed` when different frames
were produced by different models. A cache hit reuses that producer identity
instead of relabeling it as the requested routing plan. The whole-video result
cache is keyed on every input that changes the output — prompt, effective
model, sampling policy, frame count, focus window, `transcript`,
model, sampling policy, frame count, semantic analysis mode, the SHA-256
fingerprint of the normalized focus hint, focus window, `transcript`,
`audioTranscript`, and the contact-sheet flag — so changing any of those
dimensions is a cache miss, never a stale reuse. The visual dedup policy
version, threshold, and bounded candidate-frame count are also explicit in the
result-cache key and metadata; a policy change therefore cannot reuse a stale
whole-video description.
whole-video description. Result-cache v4 metadata keeps the mode and
fingerprint, never the raw user task. Guardrail metadata reports both the
requested and effective analysis modes; a requested `focused` mode without
usable user text is reported as effectively `full`.
The guardrail extracts every supported video part but describes no more than
`modalityBridgeVideoMaxVideos`. For a target proven to have
@@ -429,6 +446,7 @@ Runtime settings are DB-backed and Zod-validated:
| Key | Default | Range / behavior |
| ----------------------------------- | ----------- | --------------------------------------------------------------------------------------------------- |
| `modalityBridgeVideoEnabled` | `false` | Optional runtime, opt-in |
| `modalityBridgeVideoAnalysisMode` | `"full"` | `full` preserves generic captions; `focused` uses bounded, untrusted latest-user context |
| `modalityBridgeVideoModel` | `""` | Inherit the Vision Bridge model |
| `modalityBridgeVideoFrameCount` | `8` | 1–16 |
| `modalityBridgeVideoSamplingPolicy` | `"uniform"` | `uniform`, `scene_aware`, or proportional `segment_aware`; detector failure falls back to `uniform` |
@@ -671,7 +689,8 @@ Audio uses `modalityBridgeAudioEnabled`, `modalityBridgeAudioModel`,
`modalityBridgeCache*` settings. Audio has no legacy-key fallback because these
keys were introduced with the Modality Bridge schema.
Video uses `modalityBridgeVideoEnabled`, `modalityBridgeVideoModel`,
Video uses `modalityBridgeVideoEnabled`, `modalityBridgeVideoAnalysisMode`,
`modalityBridgeVideoModel`,
`modalityBridgeVideoFrameCount`, `modalityBridgeVideoSamplingPolicy`,
`modalityBridgeVideoMaxVideos`, and
`modalityBridgeVideoTimeout`, plus the shared `modalityBridgeCache*` settings.