[Bug]: Semantic cache — RediSearch vector search fails ("Syntax error") for TAG values containing ":" (escapeSearchValue omits ":" and other special chars) #5333

Prerequisites

• I have searched existing issues and discussions to avoid duplicates
• I am using the latest version (or have tested against main/nightly)

Description

When the semantic cache runs a vector search and a metadata TAG filter value contains a colon : — e.g. the model name (cache_by_model: true) for Ollama/vLLM names like gemma31b-q6:latest — the FT.SEARCH is rejected and the plugin logs semantic search skipped. Semantic matching then silently never fires; every request falls through and is stored fresh. Direct (hash) caching still works, so it's easy to miss.

Root cause: framework/vectorstore/redis.go → escapeSearchValue() escapes a subset of RediSearch specials but omits : (and ; / < > = + ).
For a TAG query @field:{value} at DIALECT 2, an unescaped : is a syntax error.

Model IDs commonly contain : (:latest, :q6_K) and / (provider-prefixed).

Steps to reproduce

FT.CREATE testidx PREFIX 1 td: SCHEMA model TAG
HSET td:1 model "gemma31b-q6:latest"
FT.SEARCH testidx "@model:{gemma31b-q6:latest}" DIALECT 2 # current output -> Syntax error near gemma31b-q6
FT.SEARCH testidx "@model:{gemma31b-q6:latest}" DIALECT 2 # colon escaped -> 1) td:1 (works)

Expected behavior

semantic search runs; similar prompt returns a cached (semantic) hit.

Actual behavior

native vector search failed: Syntax error at offset N near ; semantic search skipped; never a semantic hit for such model/tag values.

Affected area(s)

Framework

Version

framework v1.5.1 (built from source @ c0909f9; gateway ~v2.1.28)

Environment

- OS:              Ubuntu 24.04.4 LTS · Linux x86_64 · kernel 6.8.0-111-generic
- Go toolchain:    1.26.4  (build image golang:1.26.4-alpine3.23)
- Node:            25      (UI build image node:25-alpine3.23)
- Bifrost:         self-hosted OSS, built from source at commit c0909f9
- Vector store:    Redis 7.4.7 + RediSearch 2.10.20
                   (redis/redis-stack-server:latest), query DIALECT 2
- Deployment:      Docker, --network host, source_of_truth: split
- Embedding:       local OpenAI-compatible endpoint (Qwen3-Embedding-0.6B llamafile)

Relevant logs/output

Regression?

No response

Severity

Medium (some functionality impaired)