Add support for a local model (Qwen3)

This commit is contained in:
James Babcock committed 2026-08-09 23:52:21 -07:00
1 parent 6103c00ca3
commit 6f45888627
16 files changed
+1532 -15

No files matched your search

+11 -4
View File
@@ -1,4 +1,5 @@
import { streamAnthropicMessage } from "../native/anthropic";
import { streamLocalQwen } from "../native/llama";
import { streamOpenAiResponse } from "../native/openai";
import type {
LlmContentBlock,
@@ -20,6 +21,8 @@ import type { AssistantTurnCallbacks, AssistantTurnHandle } from "./types";
/** Safety caps so a tool-thrashing model can't loop or run forever. */
const MAX_ITERATIONS = 8;
/** The on-phone model is slower per step and unreliable on long tool chains. */
const MAX_ITERATIONS_LOCAL = 4;
const MAX_TURN_MS = 2 * 60 * 1000;
export type DirectTurnOptions = {
@@ -54,9 +57,10 @@ export class DirectAssistantBackend {
options.callbacks.onError(message);
};
const maxIterations = options.provider === "local" ? MAX_ITERATIONS_LOCAL : MAX_ITERATIONS;
const runIteration = (iteration: number): void => {
if (cancelled) return;
if (iteration >= MAX_ITERATIONS) {
if (iteration >= maxIterations) {
finishError("Assistant stopped: too many tool steps");
return;
}
@@ -115,9 +119,12 @@ export class DirectAssistantBackend {
finishError(message);
},
};
streamHandle = options.provider === "openai"
? streamOpenAiResponse(streamOptions)
: streamAnthropicMessage(streamOptions);
streamHandle =
options.provider === "openai"
? streamOpenAiResponse(streamOptions)
: options.provider === "local"
? streamLocalQwen(streamOptions)
: streamAnthropicMessage(streamOptions);
};
runIteration(0);
+145
View File
@@ -0,0 +1,145 @@
import type { LlmToolDefinition } from "./llm-protocol";
/**
* Generate a GBNF grammar (llama.cpp grammar dialect) that constrains a
* Qwen-style tool call to be well-formed: one <tool_call> block whose JSON
* names a real tool and whose arguments match that tool's schema. Used as a
* *lazy* grammar — sampling is unconstrained until the model emits
* <tool_call>, so plain text answers are unaffected, and once the block
* closes the grammar is complete (the model can only end the message, which
* serializes multi-call turns into one call per loop iteration — deliberate:
* small models are far more reliable one call at a time).
*
* Our tool schemas are flat objects of string/number/boolean/enum properties.
* Anything fancier degrades to a generic JSON value rule rather than failing.
*/
type JsonSchema = {
type?: string;
enum?: unknown[];
properties?: Record<string, JsonSchema>;
required?: string[];
description?: string;
};
/** Longest run of optional properties we'll enumerate subsets for. */
const MAX_SUBSET_PROPS = 5;
export function buildToolCallGrammar(tools: LlmToolDefinition[]): string | null {
if (!tools.length) return null;
const used = new Set<string>();
const toolRules: string[] = [];
for (let i = 0; i < tools.length; i++) {
const args = argsRule(tools[i]!.input_schema as JsonSchema, used);
toolRules.push(
`tc${i} ::= "<tool_call>" nl "{" sp ${jsonKey("name")} sp ":" sp "\\"${gbnfEscape(
tools[i]!.name,
)}\\"" sp "," sp ${jsonKey("arguments")} sp ":" sp ${args} sp "}" nl "</tool_call>"`,
);
}
used.add("sp");
used.add("nl");
const lines = [`root ::= ${toolRules.map((_, i) => `tc${i}`).join(" | ")}`, ...toolRules];
for (const [name, rule] of SHARED_RULES) {
if (used.has(name)) lines.push(rule);
}
return lines.join("\n");
}
/** Build the arguments-object expression for one tool schema. */
function argsRule(schema: JsonSchema | undefined, used: Set<string>): string {
const properties = schema?.properties ?? {};
const names = Object.keys(properties);
if (!names.length) return `"{" sp "}"`;
const required = new Set(schema?.required ?? []);
const optionalCount = names.filter((n) => !required.has(n)).length;
if (names.length > MAX_SUBSET_PROPS && optionalCount > 0) {
used.add("jobj");
return "jobj";
}
// Enumerate every subset of properties that includes all required ones,
// keeping declaration order. Small N keeps this tiny (<= 2^5 alternatives).
const members = names.map((name) => {
const value = valueExpr(properties[name]!, used);
return `${jsonKey(name)} sp ":" sp ${value}`;
});
const alternatives: string[] = [];
for (let mask = 0; mask < 1 << names.length; mask++) {
const chosen: string[] = [];
let valid = true;
for (let j = 0; j < names.length; j++) {
const included = (mask & (1 << j)) !== 0;
if (included) chosen.push(members[j]!);
else if (required.has(names[j]!)) valid = false;
}
if (!valid) continue;
if (!chosen.length) {
alternatives.push(`"{" sp "}"`);
} else {
alternatives.push(`"{" sp ${chosen.join(` sp "," sp `)} sp "}"`);
}
}
return alternatives.length === 1 ? alternatives[0]! : `(${alternatives.join(" | ")})`;
}
/** Grammar expression for one property value. */
function valueExpr(schema: JsonSchema, used: Set<string>): string {
if (Array.isArray(schema.enum) && schema.enum.length) {
if (schema.enum.every((v) => typeof v === "string")) {
return `(${schema.enum.map((v) => `"\\"${gbnfEscape(String(v))}\\""`).join(" | ")})`;
}
if (schema.enum.every((v) => typeof v === "number" || typeof v === "boolean")) {
return `(${schema.enum.map((v) => `"${String(v)}"`).join(" | ")})`;
}
}
switch (schema.type) {
case "string":
used.add("jstring");
used.add("jchar");
used.add("jhex");
return "jstring";
case "number":
used.add("jnumber");
return "jnumber";
case "integer":
used.add("jint");
return "jint";
case "boolean":
used.add("jbool");
return "jbool";
default:
for (const name of ["jvalue", "jstring", "jchar", "jhex", "jnumber", "jbool", "jarray", "jobj", "jmember"]) {
used.add(name);
}
return "jvalue";
}
}
/** A JSON object key as a GBNF literal: "name" -> "\"name\"". */
function jsonKey(name: string): string {
return `"\\"${gbnfEscape(name)}\\""`;
}
/** Escape for the inside of a GBNF double-quoted literal. */
function gbnfEscape(text: string): string {
return text.replace(/\\/g, "\\\\").replace(/"/g, '\\"').replace(/\n/g, "\\n");
}
// Terminal rules, emitted only when referenced. jchar mirrors llama.cpp's
// json.gbnf string character rule.
const SHARED_RULES: Array<[string, string]> = [
["sp", `sp ::= " "?`],
["nl", `nl ::= "\\n"`],
["jstring", `jstring ::= "\\"" jchar* "\\""`],
["jchar", `jchar ::= [^"\\\\\\x7F\\x00-\\x1F] | "\\\\" (["\\\\bfnrt] | "u" jhex jhex jhex jhex)`],
["jhex", `jhex ::= [0-9a-fA-F]`],
["jnumber", `jnumber ::= "-"? ("0" | [1-9] [0-9]*) ("." [0-9]+)? (("e" | "E") ("-" | "+")? [0-9]+)?`],
["jint", `jint ::= "-"? ("0" | [1-9] [0-9]*)`],
["jbool", `jbool ::= "true" | "false"`],
["jvalue", `jvalue ::= jstring | jnumber | jbool | "null" | jarray | jobj`],
["jarray", `jarray ::= "[" sp (jvalue (sp "," sp jvalue)*)? sp "]"`],
["jobj", `jobj ::= "{" sp (jmember (sp "," sp jmember)*)? sp "}"`],
["jmember", `jmember ::= jstring sp ":" sp jvalue`],
];
+81
View File
@@ -0,0 +1,81 @@
import type { LlmContentBlock, LlmMessage, LlmToolDefinition } from "./llm-protocol";
/**
* Render our provider-neutral message history into a Qwen3 (ChatML) prompt
* string, matching the model's chat template closely enough that its
* tool-calling priors apply: Hermes-style <tools> declarations in the system
* block, <tool_call> JSON blocks in assistant turns, and tool results wrapped
* in <tool_response> inside a user turn. The caller appends nothing — the
* returned string ends with the assistant generation header.
*/
const IM_START = "<|im_start|>";
const IM_END = "<|im_end|>";
export function buildQwenPrompt(
system: string | undefined,
messages: LlmMessage[],
tools: LlmToolDefinition[] | undefined,
): string {
const parts: string[] = [];
parts.push(`${IM_START}system\n${systemBlock(system, tools)}${IM_END}\n`);
for (const message of messages) {
if (message.role === "assistant") {
parts.push(`${IM_START}assistant\n${assistantBlock(message.content)}${IM_END}\n`);
} else {
parts.push(`${IM_START}user\n${userBlock(message.content)}${IM_END}\n`);
}
}
parts.push(`${IM_START}assistant\n`);
return parts.join("");
}
function systemBlock(system: string | undefined, tools: LlmToolDefinition[] | undefined): string {
const base = system?.trim() ?? "You are a helpful assistant.";
if (!tools?.length) return base;
const declarations = tools
.map((tool) =>
JSON.stringify({
type: "function",
function: { name: tool.name, description: tool.description, parameters: tool.input_schema },
}),
)
.join("\n");
return (
`${base}\n\n# Tools\n\n` +
`You may call one or more functions to assist with the user query.\n\n` +
`You are provided with function signatures within <tools></tools> XML tags:\n` +
`<tools>\n${declarations}\n</tools>\n\n` +
`For each function call, return a json object with function name and arguments ` +
`within <tool_call></tool_call> XML tags:\n` +
`<tool_call>\n{"name": <function-name>, "arguments": <args-json-object>}\n</tool_call>`
);
}
function assistantBlock(content: string | LlmContentBlock[]): string {
if (typeof content === "string") return content;
const parts: string[] = [];
for (const block of content) {
if (block.type === "text") {
if (block.text.trim()) parts.push(block.text.trim());
} else if (block.type === "tool_use") {
parts.push(`<tool_call>\n${JSON.stringify({ name: block.name, arguments: block.input ?? {} })}\n</tool_call>`);
}
// provider_item blocks are OpenAI-specific continuation state; skip.
}
return parts.join("\n");
}
function userBlock(content: string | LlmContentBlock[]): string {
if (typeof content === "string") return content;
const parts: string[] = [];
for (const block of content) {
if (block.type === "text") {
if (block.text.trim()) parts.push(block.text.trim());
} else if (block.type === "tool_result") {
const body = block.is_error ? `Error: ${block.content}` : block.content || "(no output)";
parts.push(`<tool_response>\n${body}\n</tool_response>`);
}
}
return parts.join("\n");
}
+16 -4
View File
@@ -1,4 +1,8 @@
import { buildAssistantSystemPrompt } from "../prompts";
import {
ASSISTANT_SYSTEM_PROMPT_BASE,
buildAssistantSystemPrompt,
describeAssistantContext,
} from "../prompts";
import { assistantBridge } from "./bridge-client";
import { DirectAssistantBackend } from "./direct-backend";
import type { LlmMessage, LlmToolDefinition } from "./llm-protocol";
@@ -108,15 +112,23 @@ export class AssistantSession {
return;
}
this.messages.push({ role: "user", content: text });
this.trimHistory();
const llm = this.config.llm;
// Local provider: keep the system prompt (and thus the tool declarations
// that follow it in the rendered prompt) byte-stable so the on-phone
// model's KV prefix cache survives across turns; the volatile context
// (time, foreground app, battery) rides along with the utterance instead.
const isLocal = llm.provider === "local";
this.messages.push({
role: "user",
content: isLocal ? `${text}\n\n(${describeAssistantContext(ctx)})` : text,
});
this.trimHistory();
this.turnHandle = this.directBackend.runTurn({
provider: llm.provider,
apiKey: llm.apiKey,
model: llm.model,
effort: llm.effort,
system: buildAssistantSystemPrompt(ctx),
system: isLocal ? ASSISTANT_SYSTEM_PROMPT_BASE : buildAssistantSystemPrompt(ctx),
messages: this.messages,
buildTools: () => this.buildToolDefinitions(),
registry: this.registry,