Files
LocalAI/pkg/mcp/localaitools/tools_middleware.go
Richard Palethorpe 085fc53bbc fix(router): production-ready request router + auto-size batch for embedding/rerank (#10104)
* fix(router): score classifier production-readiness

Conversation trimming runs through the classifier model's chat template
and trims by exact token count, sized to the model's n_batch which is
now scaled to context so long probes can't crash the backend. Missing
chat_message templates are a hard error at router build time. Router-
facing factories (Embedder/Scorer/Reranker/TokenCounter) re-resolve
ModelConfig per call so a model installed post-startup doesn't bind a
stub Backend="" config and silently fall into the loader's auto-
iterate path.

New 'vector_store' backend trace recorded inside localVectorStore on
every Search/Insert — including the backend-load-failure path that
previously vanished into an xlog.Warn — with outcome tagging
(hit/miss/empty_store/backend_load_error/find_error/insert_error/ok).
Companion cleanup drops misleading similarity:0 and input_tokens_count:0
from non-hit and text-mode traces.

Gallery local-store-development aliases to 'local-store' so the master
image satisfies pkg/model.LocalStoreBackend lookups from the embedding
cache.

Misc: llama-cpp TokenizeString reads the correct 'prompt' JSON key
(the original bug); ModelTokenize nil-guard; non-fatal mitm proxy
startup; PII 'route_local' renamed to 'allow' with docs/UI in sync;
model-editor footer no longer eats the edit area on small screens;
several config-editor template/dropdown/section fixes.

Tests: e2e router specs (casual/code-hint + long-conversation trim),
vector_store trace specs, lazy-factory specs, gallery dev-alias
resolution, Playwright trace badge + scroll regression.

Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>

* feat(backend): auto-size batch to context for embedding and rerank models

Embedding and rerank models pool over the whole input in a single physical batch (n_ubatch). With batch left at the 512 default, the backend rejects longer inputs with "input is too large to process", silently capping a large-context embedder (e.g. 8k/32k) at 512 tokens. Size n_batch to the context for these single-pass usecases, mirroring the existing FLAG_SCORE behaviour; an explicit batch: still wins.

Extracts EffectiveContextSize/EffectiveBatchSize from grpcModelOpts so the effective decode window has one home for other callers to reuse.

Adds an e2e-aio regression test that embeds a >512-token input. The AIO embedding model is switched to nomic-embed-text-v1.5 (2048 context) because the previous granite model was capped at 512 tokens and could not exercise the larger batch.

Assisted-by: claude-code:claude-opus-4-8 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>

* fix(gallery): raise arch-router scoring output cap via parallel:64

Scoring decodes the whole prompt+candidate in a single llama_decode and
reads one logit row per candidate token. The vendored llama.cpp server
caps causal output rows at n_parallel, so the default of 1 aborts with
GGML_ASSERT(n_outputs_max <= cparams.n_outputs_max) on multi-token route
labels. Set options: [parallel:64] on both arch-router quant entries to
lift the cap; kv_unified (the grpc-server default) keeps the full context
per sequence, so this does not split the KV cache.

Assisted-by: claude-code:claude-opus-4-8 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>

---------

Signed-off-by: Richard Palethorpe <io@richiejp.com>
2026-06-12 16:21:15 +02:00

79 lines
3.5 KiB
Go

package localaitools
import (
"context"
"github.com/modelcontextprotocol/go-sdk/mcp"
)
// registerMiddlewareTools wires the routing-module admin surface for the
// MCP server. The two tools mirror what the React /app/middleware page
// exposes:
//
// - get_middleware_status: read-only aggregator. The agent can ask
// "what's filtering my requests?" and get back the active PII
// pattern set, the per-model resolved enabled/override state, and
// a placeholder for routing.
// - set_pii_pattern_action: mutating. Mutations are TRANSIENT — they
// live until process restart, when patterns reload from the YAML
// defaults. The skill prompt should warn the user about that
// before applying lasting changes.
func registerMiddlewareTools(s *mcp.Server, client LocalAIClient, opts Options) {
mcp.AddTool(s, &mcp.Tool{
Name: ToolGetMiddlewareStatus,
Description: "Aggregated routing-module status: PII pattern catalogue with current actions, per-model resolved PII state and overrides, recent event count, plus the active router models and their classifier configs. Read-only.",
}, func(ctx context.Context, _ *mcp.CallToolRequest, _ struct{}) (*mcp.CallToolResult, any, error) {
status, err := client.GetMiddlewareStatus(ctx)
if err != nil {
return errorResult(err), nil, nil
}
return jsonResult(status), nil, nil
})
mcp.AddTool(s, &mcp.Tool{
Name: ToolGetRouterDecisions,
Description: "Recent intelligent-routing decisions. Each row records which router model the client called, which candidate the classifier picked, the classifier's score and latency, and a correlation id that joins back to the usage record. Filter by correlation_id, user_id, or router_model. Read-only.",
}, func(ctx context.Context, _ *mcp.CallToolRequest, args RouterDecisionsQuery) (*mcp.CallToolResult, any, error) {
decisions, err := client.GetRouterDecisions(ctx, args)
if err != nil {
return errorResult(err), nil, nil
}
return jsonResult(decisions), nil, nil
})
if opts.DisableMutating {
return
}
mcp.AddTool(s, &mcp.Tool{
Name: ToolSetPIIPatternAction,
Description: "Change a PII pattern's action (mask|block|allow) and/or disabled state in-process. TRANSIENT: the mutation is lost on restart unless followed by persist_pii_patterns. Admin-required.",
}, func(ctx context.Context, _ *mcp.CallToolRequest, args PIIPatternActionUpdate) (*mcp.CallToolResult, any, error) {
if args.ID == "" {
return errorResultf("id is required"), nil, nil
}
if args.Action == "" && args.Disabled == nil {
return errorResultf("at least one of action (mask, block, allow) or disabled must be set"), nil, nil
}
if err := client.SetPIIPatternAction(ctx, args); err != nil {
return errorResult(err), nil, nil
}
return jsonResult(map[string]any{
"id": args.ID,
"action": args.Action,
"disabled": args.Disabled,
"persisted": false,
}), nil, nil
})
mcp.AddTool(s, &mcp.Tool{
Name: ToolPersistPIIPatterns,
Description: "Snapshot the live PII redactor's per-pattern (action, disabled) state into runtime_settings.json so it re-applies on the next process start. Pairs with set_pii_pattern_action — that one is in-process; this one persists. Admin-required.",
}, func(ctx context.Context, _ *mcp.CallToolRequest, _ struct{}) (*mcp.CallToolResult, any, error) {
if err := client.PersistPIIPatterns(ctx); err != nil {
return errorResult(err), nil, nil
}
return jsonResult(map[string]any{"persisted": true}), nil, nil
})
}