@convex-dev/agent 0.0.9 → 0.0.11-alpha.0

package/src/client/index.ts

@@ -60,7 +60,13 @@ import {
   vStorageOptions,
   vTextArgs,
 } from "../validators";
-import { RunActionCtx, RunMutationCtx, RunQueryCtx, UseApi } from "./types.js";
+import {
+  MessageDoc,
+  RunActionCtx,
+  RunMutationCtx,
+  RunQueryCtx,
+  UseApi,
+} from "./types.js";
 
 export { convexToZod, zodToConvex };
 export type { ThreadDoc, MessageDoc } from "./types.js";
@@ -72,11 +78,13 @@ export type { ThreadDoc, MessageDoc } from "./types.js";
 export type ContextOptions = {
   /**
    * Whether to include tool messages in the context.
+   * By default, tool calls and results are not included.
    */
   includeToolCalls?: boolean;
   /**
    * How many recent messages to include. These are added after the search
    * messages, and do not count against the search limit.
+   * Default: 100
    */
   recentMessages?: number;
   /**
@@ -84,21 +92,24 @@ export type ContextOptions = {
    */
   searchOptions?: {
     /**
-     * The maximum number of messages to fetch.
+     * The maximum number of messages to fetch. Default is 10.
      */
     limit: number;
     /**
-     * Whether to use text search to find messages.
+     * Whether to use text search to find messages. Default is false.
      */
     textSearch?: boolean;
     /**
-     * Whether to use vector search to find messages.
+     * Whether to use vector search to find messages. Default is false.
+     * At least one of textSearch or vectorSearch must be true.
      */
     vectorSearch?: boolean;
     /**
+     * What messages around the search results to include.
+     * Default: { before: 2, after: 1 }
+     * (two before, and one after each message found in the search)
      * Note, this is after the limit is applied.
      * By default this will quadruple the number of messages fetched.
-     * (two before, and one after each message found in the search)
      */
     messageRange?: { before: number; after: number };
   };
@@ -367,7 +378,7 @@ export class Agent<AgentTools extends ToolSet> {
   ): Promise<CoreMessage[]> {
     assert(args.userId || args.threadId, "Specify userId or threadId");
     // Fetch the latest messages from the thread
-    const contextMessages: CoreMessage[] = [];
+    const contextMessages: MessageDoc[] = [];
     let included: Set<string> | undefined;
     const opts = this.mergedContextOptions(args);
     if (opts.searchOptions?.textSearch || opts.searchOptions?.vectorSearch) {
@@ -385,16 +396,14 @@ export class Agent<AgentTools extends ToolSet> {
       );
       // TODO: track what messages we used for context
       included = new Set(searchMessages.map((m) => m._id));
-      contextMessages.push(
-        ...searchMessages.map((m) => deserializeMessage(m.message!))
-      );
+      contextMessages.push(...searchMessages);
     }
     if (args.threadId && opts.recentMessages !== 0) {
       const { page } = await ctx.runQuery(
         this.component.messages.getThreadMessages,
         {
           threadId: args.threadId,
-          isTool: args.includeToolCalls ?? false,
+          isTool: opts.includeToolCalls ? undefined : false,
           paginationOpts: {
             numItems: opts.recentMessages ?? DEFAULT_RECENT_MESSAGES,
             cursor: null,
@@ -404,13 +413,13 @@ export class Agent<AgentTools extends ToolSet> {
           statuses: ["success"],
         }
       );
-      contextMessages.push(
-        ...page
-          .filter((m) => !included?.has(m._id))
-          .map((m) => deserializeMessage(m.message!))
-      );
+      contextMessages.push(...page.filter((m) => !included?.has(m._id)));
     }
-    return contextMessages;
+    return contextMessages
+      .sort((a, b) =>
+        a.order === b.order ? a.stepOrder - b.stepOrder : a.order - b.order
+      )
+      .map((m) => deserializeMessage(m.message!));
   }
 
   async getEmbeddings(messages: CoreMessage[]) {
@@ -1284,6 +1293,10 @@ type OurStreamObjectArgs<T> = StreamObjectArgs<T> &
     "onError" | "onFinish" | "abortSignal"
   >;
 
+type ThreadOutputMetadata = GenerationOutputMetadata & {
+  messageId: string;
+};
+
 interface Thread<AgentTools extends ToolSet> {
   /**
    * The target threadId, from the startThread or continueThread initializers.
@@ -1306,7 +1319,7 @@ interface Thread<AgentTools extends ToolSet> {
       Parameters<typeof generateText<TOOLS, OUTPUT, OUTPUT_PARTIAL>>[0]
     >
   ): Promise<
-    GenerateTextResult<TOOLS & AgentTools, OUTPUT> & GenerationOutputMetadata
+    GenerateTextResult<TOOLS & AgentTools, OUTPUT> & ThreadOutputMetadata
   >;
 
   /**
@@ -1326,8 +1339,7 @@ interface Thread<AgentTools extends ToolSet> {
       Parameters<typeof streamText<TOOLS, OUTPUT, PARTIAL_OUTPUT>>[0]
     >
   ): Promise<
-    StreamTextResult<TOOLS & AgentTools, PARTIAL_OUTPUT> &
-      GenerationOutputMetadata
+    StreamTextResult<TOOLS & AgentTools, PARTIAL_OUTPUT> & ThreadOutputMetadata
   >;
   /**
    * This behaves like {@link generateObject} from the "ai" package except that
@@ -1341,7 +1353,7 @@ interface Thread<AgentTools extends ToolSet> {
    */
   generateObject<T>(
     args: OurObjectArgs<T>
-  ): Promise<GenerateObjectResult<T> & GenerationOutputMetadata>;
+  ): Promise<GenerateObjectResult<T> & ThreadOutputMetadata>;
   /**
    * This behaves like {@link generateObject} from the "ai" package except that
    * it add context based on the userId and threadId and saves the input and
@@ -1354,7 +1366,7 @@ interface Thread<AgentTools extends ToolSet> {
    */
   generateObject(
     args: GenerateObjectNoSchemaOptions
-  ): Promise<GenerateObjectResult<JSONValue> & GenerationOutputMetadata>;
+  ): Promise<GenerateObjectResult<JSONValue> & ThreadOutputMetadata>;
   /**
    * This behaves like {@link streamObject} from the "ai" package except that
    * it add context based on the userId and threadId and saves the input and
@@ -1368,6 +1380,6 @@ interface Thread<AgentTools extends ToolSet> {
   streamObject<T>(
     args: OurStreamObjectArgs<T>
   ): Promise<
-    StreamObjectResult<DeepPartial<T>, T, never> & GenerationOutputMetadata
+    StreamObjectResult<DeepPartial<T>, T, never> & ThreadOutputMetadata
   >;
 }

package/src/component/messages.ts

@@ -772,7 +772,10 @@ export const _fetchVectorMessages = internalQuery({
       )
     ).filter(
       (m): m is Doc<"messages"> =>
-        m !== undefined && m !== null && (!parent || m.order <= parent.order)
+        m !== undefined &&
+        m !== null &&
+        !m.tool &&
+        (!parent || m.order <= parent.order)
     );
     messages.push(...(args.textSearchMessages ?? []));
     // TODO: prioritize more recent messages
@@ -873,6 +876,8 @@ export const textSearch = query({
           ? q.search("text", args.text).eq("userId", args.userId)
           : q.search("text", args.text).eq("threadId", args.threadId!)
       )
+      // Just in case tool messages slip through
+      .filter((q) => q.eq(q.field("tool"), false))
       .take(args.limit);
     return messages;
   },

package/src/mapping.ts

@@ -86,8 +86,8 @@ export function serializeNewMessagesInStep<TOOLS extends ToolSet>(
   // ref: https://github.com/vercel/ai/blob/main/packages/ai/core/generate-text/to-response-messages.ts
   const messages: MessageWithFileAndId[] = (
     step.toolResults.length > 0
-      ? step.response.messages.slice(0, -2)
-      : step.response.messages.slice(0, -1)
+      ? step.response.messages.slice(-2)
+      : step.response.messages.slice(-1)
   ).map(serializeMessageWithId);
   return messages;
 }
@@ -193,14 +193,10 @@ function deserializeUrl(urlOrString: string | ArrayBuffer): URL | DataContent {
 }
 
 export function promptOrMessagesToCoreMessages(args: {
-  system?: string;
   prompt?: string;
   messages?: CoreMessage[] | AIMessageWithoutId[];
 }): CoreMessage[] {
   const messages: CoreMessage[] = [];
-  if (args.system) {
-    messages.push({ role: "system", content: args.system });
-  }
   assert(args.prompt || args.messages, "messages or prompt is required");
   if (args.messages) {
     if (