@dbx-tools/appkit-mastra 0.9.176 → 0.9.179
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +14 -11
- package/index.ts +4 -4
- package/lib/index.d.ts +4 -4
- package/lib/index.js +5 -5
- package/lib/src/agents.js +12 -2
- package/lib/src/chat.d.ts +20 -0
- package/lib/src/chat.js +226 -39
- package/lib/src/config.d.ts +7 -5
- package/lib/src/config.js +3 -3
- package/lib/src/feedback-mode.d.ts +3 -3
- package/lib/src/feedback-mode.js +9 -10
- package/lib/src/follow-up-suggestions.d.ts +1 -1
- package/lib/src/follow-up-suggestions.js +28 -76
- package/lib/src/memory.d.ts +4 -1
- package/lib/src/memory.js +1 -16
- package/lib/src/model.d.ts +11 -6
- package/lib/src/model.js +56 -26
- package/lib/src/plugin.d.ts +7 -5
- package/lib/src/plugin.js +19 -17
- package/lib/src/server.d.ts +3 -1
- package/lib/src/server.js +3 -3
- package/lib/src/summarize.d.ts +10 -14
- package/lib/src/summarize.js +44 -15
- package/package.json +18 -18
- package/src/agents.ts +11 -2
- package/src/chat.ts +262 -39
- package/src/config.ts +9 -7
- package/src/feedback-mode.ts +13 -9
- package/src/follow-up-suggestions.ts +32 -80
- package/src/memory.ts +1 -16
- package/src/model.ts +101 -43
- package/src/plugin.ts +25 -18
- package/src/server.ts +2 -1
- package/src/summarize.ts +46 -14
package/src/model.ts
CHANGED
|
@@ -24,11 +24,7 @@
|
|
|
24
24
|
|
|
25
25
|
import { createWorkspaceClient, type WorkspaceClient } from "@databricks/appkit";
|
|
26
26
|
import { tryGetExecutionContext } from "@dbx-tools/appkit";
|
|
27
|
-
import {
|
|
28
|
-
SmartRouterService,
|
|
29
|
-
SMART_ROUTER_DEFAULT_SETTINGS,
|
|
30
|
-
smartRouterCandidates,
|
|
31
|
-
} from "@dbx-tools/appkit-model-gateway";
|
|
27
|
+
import { SmartRouterService, smartRouterCandidates } from "@dbx-tools/appkit-model-gateway";
|
|
32
28
|
import {
|
|
33
29
|
classes,
|
|
34
30
|
endpointCapabilities,
|
|
@@ -46,7 +42,12 @@ import { contracts, type ServingEndpointSummary } from "@dbx-tools/shared-model"
|
|
|
46
42
|
import { SMART_ROUTER_MODEL_ID, isSmartRouterModelId } from "@dbx-tools/shared-model-gateway";
|
|
47
43
|
import { RequestContext } from "@mastra/core/request-context";
|
|
48
44
|
|
|
49
|
-
import {
|
|
45
|
+
import {
|
|
46
|
+
executionContextUserId,
|
|
47
|
+
MASTRA_USER_KEY,
|
|
48
|
+
type MastraPluginConfig,
|
|
49
|
+
type User,
|
|
50
|
+
} from "./config.ts";
|
|
50
51
|
import {
|
|
51
52
|
MASTRA_MODEL_OVERRIDE_KEY,
|
|
52
53
|
MASTRA_RESOLVED_MODEL_KEY,
|
|
@@ -56,9 +57,16 @@ import {
|
|
|
56
57
|
import { recordActiveTraceAuth, recordActiveTraceModel } from "./telemetry.ts";
|
|
57
58
|
|
|
58
59
|
type ModelClass = contracts.ModelClass;
|
|
60
|
+
type ModelOptimization = contracts.ModelOptimization;
|
|
59
61
|
const { parseModelClass } = classes;
|
|
60
62
|
const { selectModel } = resolve;
|
|
61
|
-
const smartRouter = new SmartRouterService(
|
|
63
|
+
const smartRouter = new SmartRouterService();
|
|
64
|
+
const MASTRA_SMART_ROUTER_RESOLUTION_KEY = "mastra__smart_router_resolution";
|
|
65
|
+
|
|
66
|
+
interface SmartRouterModelResolution {
|
|
67
|
+
readonly modelId: string;
|
|
68
|
+
readonly source: string;
|
|
69
|
+
}
|
|
62
70
|
|
|
63
71
|
/** Virtual endpoint shown when Smart Router is enabled for the active identity. */
|
|
64
72
|
export const SMART_ROUTER_ENDPOINT: ServingEndpointSummary = {
|
|
@@ -90,14 +98,18 @@ export interface BuildModelOverrides {
|
|
|
90
98
|
*/
|
|
91
99
|
modelId?: string;
|
|
92
100
|
/**
|
|
93
|
-
* Chat capability class
|
|
94
|
-
* supplied
|
|
95
|
-
*
|
|
96
|
-
*
|
|
97
|
-
* available model in the class
|
|
98
|
-
*
|
|
101
|
+
* Chat capability class for internal model work. When no model id is
|
|
102
|
+
* supplied, this bypasses request and environment model overrides so
|
|
103
|
+
* background tasks such as titles and follow-ups cannot inherit the
|
|
104
|
+
* primary agent's Smart Router selection. The live catalogue chooses
|
|
105
|
+
* the top available model in the class, falling back to the class's
|
|
106
|
+
* static list when the workspace has none.
|
|
99
107
|
*/
|
|
100
108
|
modelClass?: ModelClass;
|
|
109
|
+
/** Profile dimension used to rank eligible automatic model candidates. */
|
|
110
|
+
optimizeFor?: ModelOptimization;
|
|
111
|
+
/** Require documented structured-output support for the selected endpoint. */
|
|
112
|
+
requiresStructuredOutput?: boolean;
|
|
101
113
|
/** Arbitrary request, messages, JSON, or text used to build Smart Router task context. */
|
|
102
114
|
smartRouterInput?: unknown;
|
|
103
115
|
/** Explicit workspace client for standalone callers such as CLIs. */
|
|
@@ -108,6 +120,7 @@ export interface BuildModelOverrides {
|
|
|
108
120
|
interface ResolvedModelWorkspaceClient {
|
|
109
121
|
client: WorkspaceClient;
|
|
110
122
|
auth: "obo" | "service-principal";
|
|
123
|
+
cacheIdentity?: string;
|
|
111
124
|
}
|
|
112
125
|
|
|
113
126
|
/** Resolve an OBO workspace client when present, else a service-principal client. */
|
|
@@ -120,6 +133,7 @@ function resolveModelWorkspaceContext(
|
|
|
120
133
|
return {
|
|
121
134
|
client: executionContext.client,
|
|
122
135
|
auth: "isUserContext" in executionContext ? "obo" : "service-principal",
|
|
136
|
+
cacheIdentity: executionContextUserId(executionContext),
|
|
123
137
|
};
|
|
124
138
|
}
|
|
125
139
|
return { client: createWorkspaceClient(), auth: "service-principal" };
|
|
@@ -134,7 +148,9 @@ function selectionInput(
|
|
|
134
148
|
config: MastraPluginConfig,
|
|
135
149
|
requested: string | undefined,
|
|
136
150
|
defaultClass: ModelClass | undefined,
|
|
151
|
+
cacheIdentity: string | undefined,
|
|
137
152
|
serving = resolveServingConfig(config),
|
|
153
|
+
optimization?: Pick<BuildModelOverrides, "optimizeFor" | "requiresStructuredOutput">,
|
|
138
154
|
) {
|
|
139
155
|
const requestedClass = requested !== undefined ? parseModelClass(requested) : null;
|
|
140
156
|
const explicit = requestedClass === null ? requested : undefined;
|
|
@@ -144,9 +160,14 @@ function selectionInput(
|
|
|
144
160
|
fuzzy: serving.fuzzy,
|
|
145
161
|
threshold: serving.threshold,
|
|
146
162
|
...(modelClass !== undefined ? { modelClass } : {}),
|
|
163
|
+
...(optimization?.optimizeFor !== undefined ? { optimizeFor: optimization.optimizeFor } : {}),
|
|
164
|
+
...(optimization?.requiresStructuredOutput !== undefined
|
|
165
|
+
? { requiresStructuredOutput: optimization.requiresStructuredOutput }
|
|
166
|
+
: {}),
|
|
147
167
|
fallbacks: serving.fallbacks.slice(),
|
|
148
168
|
liveOnly: requested === undefined,
|
|
149
169
|
ttlMs: serving.ttlMs,
|
|
170
|
+
...(cacheIdentity !== undefined ? { cacheIdentity } : {}),
|
|
150
171
|
};
|
|
151
172
|
}
|
|
152
173
|
|
|
@@ -166,8 +187,10 @@ export function resolveDefaultModelId(
|
|
|
166
187
|
return SMART_ROUTER_MODEL_ID;
|
|
167
188
|
}
|
|
168
189
|
const requested = overrides.modelId ?? process.env.DATABRICKS_SERVING_ENDPOINT_NAME;
|
|
169
|
-
return selection.resolveModel(
|
|
170
|
-
|
|
190
|
+
return selection.resolveModel(
|
|
191
|
+
endpoints,
|
|
192
|
+
selectionInput(config, requested, overrides.modelClass, undefined, undefined, overrides),
|
|
193
|
+
).modelId;
|
|
171
194
|
}
|
|
172
195
|
|
|
173
196
|
/**
|
|
@@ -196,7 +219,7 @@ export async function buildModel(
|
|
|
196
219
|
const resolved = overrides.workspaceClient
|
|
197
220
|
? { client: overrides.workspaceClient, auth: "service-principal" as const }
|
|
198
221
|
: resolveModelWorkspaceContext(context);
|
|
199
|
-
const { client, auth } = resolved;
|
|
222
|
+
const { client, auth, cacheIdentity } = resolved;
|
|
200
223
|
recordActiveTraceAuth(auth);
|
|
201
224
|
const clientConfig = client.config;
|
|
202
225
|
const host = (await clientConfig.getHost()).toString();
|
|
@@ -204,20 +227,29 @@ export async function buildModel(
|
|
|
204
227
|
await clientConfig.authenticate(headers);
|
|
205
228
|
const logger = log.logger(config);
|
|
206
229
|
const serving = resolveServingConfig(config);
|
|
207
|
-
const
|
|
208
|
-
|
|
209
|
-
|
|
230
|
+
const internalClassOnly = overrides.modelClass !== undefined && overrides.modelId === undefined;
|
|
231
|
+
const override =
|
|
232
|
+
serving.allowOverride && !internalClassOnly
|
|
233
|
+
? (context.get(MASTRA_MODEL_OVERRIDE_KEY) as string | undefined)
|
|
234
|
+
: undefined;
|
|
210
235
|
|
|
211
236
|
// The override / agent default / env value can be either a concrete
|
|
212
237
|
// endpoint name or a model class slug ("chat-thinking" /
|
|
213
238
|
// "chat-balanced" / "chat-fast"). A class slug becomes a class intent
|
|
214
239
|
// (let the live catalogue pick the best model in that band); anything
|
|
215
240
|
// else is an explicit name fuzzy-matched against the catalogue. An
|
|
216
|
-
// internal `overrides.modelClass` (e.g.
|
|
217
|
-
//
|
|
218
|
-
|
|
241
|
+
// internal `overrides.modelClass` (e.g. titles, follow-ups, or the chart
|
|
242
|
+
// planner) is authoritative unless that caller also pins `modelId`. It must
|
|
243
|
+
// not inherit the primary agent's request or environment selection.
|
|
244
|
+
const requested =
|
|
245
|
+
override ??
|
|
246
|
+
overrides.modelId ??
|
|
247
|
+
(!internalClassOnly ? process.env.DATABRICKS_SERVING_ENDPOINT_NAME : undefined);
|
|
219
248
|
const smartRouterInput = overrides.smartRouterInput ?? context.get(MASTRA_SMART_ROUTER_INPUT_KEY);
|
|
220
249
|
const smartRouterRequested = isSmartRouterModelId(requested);
|
|
250
|
+
if (smartRouterRequested && smartRouterInput === undefined) {
|
|
251
|
+
throw new Error("Smart Router requires task input before it can resolve a concrete model");
|
|
252
|
+
}
|
|
221
253
|
let modelId: string;
|
|
222
254
|
let source: string;
|
|
223
255
|
if (
|
|
@@ -225,36 +257,62 @@ export async function buildModel(
|
|
|
225
257
|
overrides.modelClass === undefined &&
|
|
226
258
|
smartRouterInput !== undefined
|
|
227
259
|
) {
|
|
228
|
-
|
|
229
|
-
|
|
230
|
-
|
|
231
|
-
|
|
232
|
-
|
|
233
|
-
|
|
234
|
-
|
|
235
|
-
|
|
236
|
-
|
|
237
|
-
|
|
238
|
-
|
|
239
|
-
|
|
240
|
-
|
|
241
|
-
|
|
242
|
-
|
|
243
|
-
|
|
244
|
-
|
|
245
|
-
|
|
246
|
-
|
|
247
|
-
|
|
260
|
+
let pending = context.get(MASTRA_SMART_ROUTER_RESOLUTION_KEY) as
|
|
261
|
+
Promise<SmartRouterModelResolution> | undefined;
|
|
262
|
+
if (!pending) {
|
|
263
|
+
pending = (async () => {
|
|
264
|
+
const endpoints = await listServingEndpoints(client, host, {
|
|
265
|
+
ttlMs: serving.ttlMs,
|
|
266
|
+
...(cacheIdentity !== undefined ? { cacheIdentity } : {}),
|
|
267
|
+
});
|
|
268
|
+
const candidates = smartRouterCandidates(
|
|
269
|
+
endpoints.filter((endpoint) => endpointCapabilities(endpoint).tools),
|
|
270
|
+
"codex",
|
|
271
|
+
);
|
|
272
|
+
// An explicit Smart Router request does not need a synthetic preflight:
|
|
273
|
+
// the real selection is the availability check. Auto mode still probes
|
|
274
|
+
// so it can fall back to the configured model policy when unavailable.
|
|
275
|
+
const enabled = smartRouterRequested || (await smartRouter.isEnabled(candidates, client));
|
|
276
|
+
if (enabled) {
|
|
277
|
+
const selected = await smartRouter.select(smartRouterInput, candidates, { client });
|
|
278
|
+
return { modelId: selected.value.name, source: SMART_ROUTER_MODEL_ID };
|
|
279
|
+
}
|
|
280
|
+
const fallback = await selectModel(
|
|
281
|
+
client,
|
|
282
|
+
host,
|
|
283
|
+
selectionInput(
|
|
284
|
+
config,
|
|
285
|
+
requested,
|
|
286
|
+
overrides.modelClass,
|
|
287
|
+
cacheIdentity,
|
|
288
|
+
serving,
|
|
289
|
+
overrides,
|
|
290
|
+
),
|
|
291
|
+
);
|
|
292
|
+
return { modelId: fallback.modelId, source: fallback.source };
|
|
293
|
+
})();
|
|
294
|
+
context.set(MASTRA_SMART_ROUTER_RESOLUTION_KEY, pending);
|
|
295
|
+
}
|
|
296
|
+
try {
|
|
297
|
+
({ modelId, source } = await pending);
|
|
298
|
+
} catch (error) {
|
|
299
|
+
if (context.get(MASTRA_SMART_ROUTER_RESOLUTION_KEY) === pending) {
|
|
300
|
+
context.set(MASTRA_SMART_ROUTER_RESOLUTION_KEY, undefined);
|
|
301
|
+
}
|
|
302
|
+
throw error;
|
|
248
303
|
}
|
|
249
304
|
} else {
|
|
250
305
|
const resolved = await selectModel(
|
|
251
306
|
client,
|
|
252
307
|
host,
|
|
253
|
-
selectionInput(config, requested, overrides.modelClass, serving),
|
|
308
|
+
selectionInput(config, requested, overrides.modelClass, cacheIdentity, serving, overrides),
|
|
254
309
|
);
|
|
255
310
|
modelId = resolved.modelId;
|
|
256
311
|
source = resolved.source;
|
|
257
312
|
}
|
|
313
|
+
if (isSmartRouterModelId(modelId)) {
|
|
314
|
+
throw new Error("Smart Router did not resolve to a concrete serving endpoint");
|
|
315
|
+
}
|
|
258
316
|
logger.debug("model selected", { modelId, source, requested });
|
|
259
317
|
context.set(MASTRA_RESOLVED_MODEL_KEY, modelId);
|
|
260
318
|
recordActiveTraceModel(modelId);
|
package/src/plugin.ts
CHANGED
|
@@ -82,7 +82,7 @@ import {
|
|
|
82
82
|
type GenieChatRequest,
|
|
83
83
|
type MastraClientConfig,
|
|
84
84
|
type StatementData,
|
|
85
|
-
type
|
|
85
|
+
type FeedbackDestination,
|
|
86
86
|
MastraFeedbackRequestSchema,
|
|
87
87
|
type MlflowFeedbackRequest,
|
|
88
88
|
type GenieFeedbackRequest,
|
|
@@ -122,7 +122,7 @@ import {
|
|
|
122
122
|
modelCatalogueDefaults,
|
|
123
123
|
statementDataDefaults,
|
|
124
124
|
} from "./defaults.ts";
|
|
125
|
-
import {
|
|
125
|
+
import { resolveFeedbackModes } from "./feedback-mode.ts";
|
|
126
126
|
import { resolveFollowUpSuggestionsConfig } from "./follow-up-suggestions.ts";
|
|
127
127
|
import {
|
|
128
128
|
deleteGenieThread,
|
|
@@ -274,8 +274,8 @@ export class MastraPlugin extends Plugin<MastraPluginConfig> {
|
|
|
274
274
|
* {@link buildAgentAndServer} and registered on the Mastra instance.
|
|
275
275
|
*/
|
|
276
276
|
private mcp: ResolvedMcp | null = null;
|
|
277
|
-
/** Resolved feedback
|
|
278
|
-
private
|
|
277
|
+
/** Resolved feedback destinations published to server and browser clients. */
|
|
278
|
+
private resolvedFeedbackModes: FeedbackDestination[] = [];
|
|
279
279
|
/**
|
|
280
280
|
* Resolved AppKit identity mode: the identity the chat path's Databricks
|
|
281
281
|
* calls run as. Read once at setup so an invalid `genieIdentity` fails the app
|
|
@@ -287,11 +287,10 @@ export class MastraPlugin extends Plugin<MastraPluginConfig> {
|
|
|
287
287
|
// app boot with a clear message rather than 500-ing the first chat request.
|
|
288
288
|
this.identityMode = resolveIdentityMode(this.config.genieIdentity);
|
|
289
289
|
const spaces = resolveGenieSpaces(this.config, this.context);
|
|
290
|
-
this.
|
|
290
|
+
this.resolvedFeedbackModes = resolveFeedbackModes(this.config.feedbackMode, {
|
|
291
291
|
genie: this.config.genieAgentMode !== false && Object.keys(spaces).length > 0,
|
|
292
292
|
mlflow: mlflowEnabled(),
|
|
293
293
|
});
|
|
294
|
-
this.config.feedbackMode = this.resolvedFeedbackMode;
|
|
295
294
|
if (this.config.observability !== false) {
|
|
296
295
|
await configureNativeHttpTelemetry({
|
|
297
296
|
// A low-level request without an active semantic span becomes a noisy
|
|
@@ -497,7 +496,7 @@ export class MastraPlugin extends Plugin<MastraPluginConfig> {
|
|
|
497
496
|
basePath: this.basePath,
|
|
498
497
|
defaultAgent: this.built?.defaultAgentId ?? FALLBACK_AGENT_ID,
|
|
499
498
|
agents: Object.keys(this.built?.agents ?? {}),
|
|
500
|
-
|
|
499
|
+
feedbackModes: this.feedbackModes(),
|
|
501
500
|
backgroundTurns: this.config.backgroundTurns !== false,
|
|
502
501
|
followUpSuggestions: followUps?.mode ?? "none",
|
|
503
502
|
followUpSuggestionLimit: followUps?.limit ?? 3,
|
|
@@ -509,9 +508,14 @@ export class MastraPlugin extends Plugin<MastraPluginConfig> {
|
|
|
509
508
|
return config as unknown as Record<string, unknown>;
|
|
510
509
|
}
|
|
511
510
|
|
|
512
|
-
/** Resolve the
|
|
513
|
-
private
|
|
514
|
-
return this.
|
|
511
|
+
/** Resolve the feedback destinations exposed for per-message routing. */
|
|
512
|
+
private feedbackModes(): FeedbackDestination[] {
|
|
513
|
+
return this.resolvedFeedbackModes;
|
|
514
|
+
}
|
|
515
|
+
|
|
516
|
+
/** Whether one feedback destination is enabled. */
|
|
517
|
+
private feedbackEnabled(mode: FeedbackDestination): boolean {
|
|
518
|
+
return this.resolvedFeedbackModes.includes(mode);
|
|
515
519
|
}
|
|
516
520
|
|
|
517
521
|
override injectRoutes(router: IAppRouter): void {
|
|
@@ -984,15 +988,14 @@ export class MastraPlugin extends Plugin<MastraPluginConfig> {
|
|
|
984
988
|
return this.execute(() => write(client), genieConversationWriteDefaults);
|
|
985
989
|
}
|
|
986
990
|
|
|
987
|
-
/** Register
|
|
991
|
+
/** Register every feedback destination enabled by plugin mode. */
|
|
988
992
|
private registerFeedbackRoutes(router: IAppRouter): void {
|
|
989
993
|
this.route(router, {
|
|
990
994
|
name: "feedback",
|
|
991
995
|
method: "post",
|
|
992
996
|
path: MASTRA_ROUTES.feedback,
|
|
993
997
|
handler: async (req, res) => {
|
|
994
|
-
|
|
995
|
-
if (mode === "none") {
|
|
998
|
+
if (this.resolvedFeedbackModes.length === 0) {
|
|
996
999
|
res.status(404).json({ ok: false });
|
|
997
1000
|
return;
|
|
998
1001
|
}
|
|
@@ -1008,7 +1011,7 @@ export class MastraPlugin extends Plugin<MastraPluginConfig> {
|
|
|
1008
1011
|
});
|
|
1009
1012
|
return;
|
|
1010
1013
|
}
|
|
1011
|
-
if (parsed.data.mode
|
|
1014
|
+
if (!this.feedbackEnabled(parsed.data.mode)) {
|
|
1012
1015
|
res.status(400).json({ ok: false, error: INVALID_FEEDBACK_MESSAGE });
|
|
1013
1016
|
return;
|
|
1014
1017
|
}
|
|
@@ -1042,7 +1045,7 @@ export class MastraPlugin extends Plugin<MastraPluginConfig> {
|
|
|
1042
1045
|
method: "get",
|
|
1043
1046
|
path: MASTRA_ROUTES.mlflowExperiment,
|
|
1044
1047
|
handler: async (req, res) => {
|
|
1045
|
-
if (this.
|
|
1048
|
+
if (!this.feedbackEnabled("mlflow")) {
|
|
1046
1049
|
res.json({ url: null });
|
|
1047
1050
|
return;
|
|
1048
1051
|
}
|
|
@@ -1454,7 +1457,10 @@ export class MastraPlugin extends Plugin<MastraPluginConfig> {
|
|
|
1454
1457
|
identity: (request) =>
|
|
1455
1458
|
useServicePrincipal(this.identityMode, request) ? "service-principal" : "obo",
|
|
1456
1459
|
});
|
|
1457
|
-
this.mastraServer = new MastraServer(
|
|
1460
|
+
this.mastraServer = new MastraServer(
|
|
1461
|
+
this.config,
|
|
1462
|
+
{ mlflowFeedback: this.feedbackEnabled("mlflow") },
|
|
1463
|
+
{
|
|
1458
1464
|
app: this.mastraApp,
|
|
1459
1465
|
mastra: this.mastra,
|
|
1460
1466
|
prefix: "",
|
|
@@ -1465,7 +1471,8 @@ export class MastraPlugin extends Plugin<MastraPluginConfig> {
|
|
|
1465
1471
|
version: "v7",
|
|
1466
1472
|
}),
|
|
1467
1473
|
],
|
|
1468
|
-
|
|
1474
|
+
},
|
|
1475
|
+
);
|
|
1469
1476
|
await this.mastraServer.init();
|
|
1470
1477
|
attachStaleResumeRecovery(this.mastraApp);
|
|
1471
1478
|
this.logger.info("ready", {
|
|
@@ -1476,7 +1483,7 @@ export class MastraPlugin extends Plugin<MastraPluginConfig> {
|
|
|
1476
1483
|
apiAccess: this.config.apiAccess ?? "scoped",
|
|
1477
1484
|
mcp: this.mcp ? `${this.basePath}${this.mcp.httpPath}` : "off",
|
|
1478
1485
|
lakebase: memoryPool !== undefined,
|
|
1479
|
-
feedback: this.
|
|
1486
|
+
feedback: this.feedbackModes(),
|
|
1480
1487
|
observability: observability !== undefined ? "mlflow" : "off",
|
|
1481
1488
|
});
|
|
1482
1489
|
}
|
package/src/server.ts
CHANGED
|
@@ -232,11 +232,12 @@ export class MastraServer extends ExpressMastraServer {
|
|
|
232
232
|
|
|
233
233
|
constructor(
|
|
234
234
|
private config: MastraPluginConfig,
|
|
235
|
+
feedback: { mlflowFeedback: boolean },
|
|
235
236
|
...args: ConstructorParameters<typeof ExpressMastraServer>
|
|
236
237
|
) {
|
|
237
238
|
super(...args);
|
|
238
239
|
this.log = log.logger(config);
|
|
239
|
-
this.mlflowFeedback =
|
|
240
|
+
this.mlflowFeedback = feedback.mlflowFeedback;
|
|
240
241
|
}
|
|
241
242
|
|
|
242
243
|
override registerAuthMiddleware(): void {
|
package/src/summarize.ts
CHANGED
|
@@ -1,16 +1,9 @@
|
|
|
1
1
|
/**
|
|
2
2
|
* Small-tier summarization for the Mastra plugin.
|
|
3
3
|
*
|
|
4
|
-
*
|
|
4
|
+
* A dedicated `summarize` tool backed by the fast / small chat tier
|
|
5
5
|
* ({@link model.ModelClass.ChatFast}) resolved through the same
|
|
6
|
-
* `/serving-endpoints` pipeline as the main agents
|
|
7
|
-
*
|
|
8
|
-
* - A dedicated `summarize` tool (see {@link buildSummarizeTool})
|
|
9
|
-
* agents can call to condense arbitrary text without burning the
|
|
10
|
-
* heavyweight chat model.
|
|
11
|
-
* - The model + instructions Mastra's memory uses to auto-name
|
|
12
|
-
* conversation threads (`generateTitle`), so titling reuses the
|
|
13
|
-
* same small tier rather than the agent's primary model.
|
|
6
|
+
* `/serving-endpoints` pipeline as the main agents.
|
|
14
7
|
*
|
|
15
8
|
* Mirrors the chart-planner wiring in `chart.ts`: a per-config cached
|
|
16
9
|
* `Agent` on the fast tier, invoked via `agent.generate(...)` inside
|
|
@@ -25,6 +18,7 @@ import { Agent } from "@mastra/core/agent";
|
|
|
25
18
|
import type { MastraModelConfig } from "@mastra/core/llm";
|
|
26
19
|
import type { RequestContext } from "@mastra/core/request-context";
|
|
27
20
|
import { createTool } from "@mastra/core/tools";
|
|
21
|
+
import { type LanguageModelMiddleware, wrapLanguageModel } from "ai";
|
|
28
22
|
import { z } from "zod";
|
|
29
23
|
|
|
30
24
|
import type { MastraPluginConfig } from "./config.ts";
|
|
@@ -33,6 +27,8 @@ import { TYPOGRAPHY_RULE } from "./style.ts";
|
|
|
33
27
|
|
|
34
28
|
/** Fast / small chat tier used for both titling and summaries. */
|
|
35
29
|
const SUMMARY_MODEL_CLASS = ModelClass.ChatFast;
|
|
30
|
+
const TITLE_INPUT_CHAR_LIMIT = 4_000;
|
|
31
|
+
const TITLE_OUTPUT_TOKEN_LIMIT = 32;
|
|
36
32
|
|
|
37
33
|
/** System prompt for the summarizer agent (and the `summarize` tool). */
|
|
38
34
|
const SUMMARIZER_INSTRUCTIONS = [
|
|
@@ -46,9 +42,9 @@ const SUMMARIZER_INSTRUCTIONS = [
|
|
|
46
42
|
].join("\n");
|
|
47
43
|
|
|
48
44
|
/**
|
|
49
|
-
* Instructions Mastra's `generateTitle` hands the
|
|
50
|
-
*
|
|
51
|
-
*
|
|
45
|
+
* Instructions Mastra's `generateTitle` hands the agent model to name a
|
|
46
|
+
* conversation thread from its opening turn. Kept terse so the model returns
|
|
47
|
+
* a bare title with no decoration.
|
|
52
48
|
*/
|
|
53
49
|
export const TITLE_INSTRUCTIONS = [
|
|
54
50
|
"Generate a short, specific title for this conversation, 3 to 6 words.",
|
|
@@ -59,8 +55,7 @@ export const TITLE_INSTRUCTIONS = [
|
|
|
59
55
|
].join(" ");
|
|
60
56
|
|
|
61
57
|
/**
|
|
62
|
-
* Resolve the small-tier model for summarization
|
|
63
|
-
* both the summarizer agent and Mastra memory's `generateTitle`.
|
|
58
|
+
* Resolve the small-tier model for summarization.
|
|
64
59
|
*
|
|
65
60
|
* Returned as a `requestContext`-taking function (a Mastra
|
|
66
61
|
* `DynamicArgument<MastraModelConfig>`) so each call mints user-scoped
|
|
@@ -73,6 +68,43 @@ export function summaryModel(
|
|
|
73
68
|
buildModel(config, requestContext, { modelClass: SUMMARY_MODEL_CLASS });
|
|
74
69
|
}
|
|
75
70
|
|
|
71
|
+
const titleModelLimits: LanguageModelMiddleware = {
|
|
72
|
+
async transformParams({ params }) {
|
|
73
|
+
let remaining = TITLE_INPUT_CHAR_LIMIT;
|
|
74
|
+
const prompt = params.prompt.map((message) => {
|
|
75
|
+
if (message.role !== "user") return message;
|
|
76
|
+
return {
|
|
77
|
+
...message,
|
|
78
|
+
content: message.content.map((part) => {
|
|
79
|
+
if (part.type !== "text") return part;
|
|
80
|
+
const text = remaining > 0 ? part.text.slice(0, remaining) : "";
|
|
81
|
+
remaining -= text.length;
|
|
82
|
+
return { ...part, text };
|
|
83
|
+
}),
|
|
84
|
+
};
|
|
85
|
+
});
|
|
86
|
+
return {
|
|
87
|
+
...params,
|
|
88
|
+
prompt,
|
|
89
|
+
maxOutputTokens: TITLE_OUTPUT_TOKEN_LIMIT,
|
|
90
|
+
};
|
|
91
|
+
},
|
|
92
|
+
};
|
|
93
|
+
|
|
94
|
+
/** Resolve a bounded, tool-free ChatFast model for automatic thread titles. */
|
|
95
|
+
export function titleModel(
|
|
96
|
+
config: MastraPluginConfig,
|
|
97
|
+
): (args: { requestContext: RequestContext }) => Promise<MastraModelConfig> {
|
|
98
|
+
return async ({ requestContext }) =>
|
|
99
|
+
wrapLanguageModel({
|
|
100
|
+
model: await buildModel(config, requestContext, {
|
|
101
|
+
modelClass: SUMMARY_MODEL_CLASS,
|
|
102
|
+
optimizeFor: "speed",
|
|
103
|
+
}),
|
|
104
|
+
middleware: titleModelLimits,
|
|
105
|
+
});
|
|
106
|
+
}
|
|
107
|
+
|
|
76
108
|
/**
|
|
77
109
|
* One summarizer `Agent` per plugin config, cached on config-object
|
|
78
110
|
* identity so a hot tool path doesn't pay the constructor cost each
|