Skip to content

feat: add remote LLM backend with averaged embedding optimization #325

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
olyashok wants to merge 2 commits into
base: main
Choose a base branch
from
Open

feat: add remote LLM backend with averaged embedding optimization #325

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c51b7fe
Add remote LLM backend support with averaged embedding optimization
Mar 7, 2026
575b9ea
Enhance embedding functionality with retry logic for transient errors
cellect-bot Mar 11, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
68 changes: 68 additions & 0 deletions server/docker-compose.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,68 @@
name: qmd

services:
qmd-embed:
image: ghcr.io/ggml-org/llama.cpp:server-cuda
container_name: qmd-embed
ports:
- "8081:8080"
volumes:
- qmd-models:/models
command: >
-m /models/embeddinggemma-300M-Q8_0.gguf
--embedding --pooling mean
--host 0.0.0.0 --port 8080
-ngl 99 -b 4096 -ub 4096 -np 4 -c 16384
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
restart: unless-stopped

qmd-rerank:
image: ghcr.io/ggml-org/llama.cpp:server-cuda
container_name: qmd-rerank
ports:
- "8082:8080"
volumes:
- qmd-models:/models
command: >
-m /models/qwen3-reranker-0.6b-q8_0.gguf
--reranking --pooling rank
--host 0.0.0.0 --port 8080
-ngl 99 -b 2048 -ub 2048 -c 4096
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
restart: unless-stopped

qmd-generate:
image: ghcr.io/ggml-org/llama.cpp:server-cuda
container_name: qmd-generate
ports:
- "8083:8080"
volumes:
- qmd-models:/models
command: >
-m /models/qmd-query-expansion-1.7B-q4_k_m.gguf
--host 0.0.0.0 --port 8080
-ngl 99 -c 2048
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
restart: unless-stopped

volumes:
qmd-models:
external: true
155 changes: 152 additions & 3 deletions src/cli/qmd.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,7 @@ import {
hashContent,
extractTitle,
formatDocForEmbedding,
chunkDocument,
chunkDocumentByTokens,
clearCache,
getCacheKey,
Expand Down Expand Up @@ -74,8 +75,15 @@ import {
generateEmbeddings,
syncConfigToDb,
type ReindexResult,
findLocalQmdDir,
initLocalQmdDir,
setCliQmdDir,
getCliQmdDir,
getEffectiveQmdDir,
setQmdDirConfigLoader,
} from "../store.js";
import { disposeDefaultLlamaCpp, getDefaultLlamaCpp, withLLMSession, pullModels, DEFAULT_EMBED_MODEL_URI, DEFAULT_GENERATE_MODEL_URI, DEFAULT_RERANK_MODEL_URI, DEFAULT_MODEL_CACHE_DIR } from "../llm.js";
import { RemoteLLM, loadRemoteConfig, saveRemoteConfig, clearRemoteConfig, isRemoteConfigured, getDefaultRemoteLLM, disposeDefaultRemoteLLM, withRemoteLLMSession, loadQmdDirConfig, saveQmdDirConfig, clearQmdDirConfig } from "../llm-remote.js";
import {
formatSearchResults,
formatDocuments,
Expand All @@ -95,13 +103,50 @@ import {
listAllContexts,
setConfigIndexName,
loadConfig,
setConfigDirResolver,
} from "../collections.js";
import { getEmbeddedQmdSkillContent, getEmbeddedQmdSkillFiles } from "../embedded-skills.js";

// Enable production mode - allows using default database path
// Tests must set INDEX_PATH or use createStore() with explicit path
enableProductionMode();

// Wire .qmd directory config resolution
setQmdDirConfigLoader(loadQmdDirConfig);
setConfigDirResolver(() => getEffectiveQmdDir());

// =============================================================================
// LLM Backend Selection
// =============================================================================

let forceLocalMode = false;

function setForceLocalMode(force: boolean): void {
forceLocalMode = force;
}

function shouldUseRemote(): boolean {
if (forceLocalMode) return false;
return isRemoteConfigured();
}

async function withLLMSessionAuto<T>(
fn: (session: any) => Promise<T>,
options?: any
): Promise<T> {
if (shouldUseRemote()) {
return withRemoteLLMSession(fn, options);
}
return withLLMSession(fn, options);
}

async function disposeAllLLM(): Promise<void> {
if (shouldUseRemote()) {
await disposeDefaultRemoteLLM();
}
await disposeDefaultLlamaCpp();
}

// =============================================================================
// Store/DB lifecycle (no legacy singletons in store.ts)
// =============================================================================
Expand Down Expand Up @@ -2125,12 +2170,13 @@ async function vectorSearch(query: string, opts: OutputOptions, _model: string =

checkIndexHealth(store.db);

await withLLMSession(async () => {
await withLLMSessionAuto(async () => {
let results = await vectorSearchQuery(store, query, {
collection: singleCollection,
limit: opts.all ? 500 : (opts.limit || 10),
minScore: opts.minScore || 0.3,
intent: opts.intent,
llm: shouldUseRemote() ? getDefaultRemoteLLM() : undefined,
hooks: {
onExpand: (original, expanded) => {
logExpansionTree(original, expanded);
Expand Down Expand Up @@ -2181,7 +2227,7 @@ async function querySearch(query: string, opts: OutputOptions, _embedModel: stri
// Intent can come from --intent flag or from intent: line in query document
const intent = opts.intent || parsed?.intent;

await withLLMSession(async () => {
await withLLMSessionAuto(async () => {
let results;

if (parsed) {
Expand Down Expand Up @@ -2234,6 +2280,7 @@ async function querySearch(query: string, opts: OutputOptions, _embedModel: stri
candidateLimit: opts.candidateLimit,
explain: !!opts.explain,
intent,
llm: shouldUseRemote() ? getDefaultRemoteLLM() : undefined,
hooks: {
onStrongSignal: (score) => {
process.stderr.write(`${c.dim}Strong BM25 signal (${score.toFixed(2)}) — skipping expansion${c.reset}\n`);
Expand Down Expand Up @@ -2349,6 +2396,12 @@ function parseCLI() {
http: { type: "boolean" },
daemon: { type: "boolean" },
port: { type: "string" },
// Remote LLM options
local: { type: "boolean" },
"embed-url": { type: "string" },
"rerank-url": { type: "string" },
"generate-url": { type: "string" },
"generate-model": { type: "string" },
},
allowPositionals: true,
strict: false, // Allow unknown options to pass through
Expand Down Expand Up @@ -2543,6 +2596,11 @@ function showHelp(): void {
console.log(" qmd context add/list/rm - Attach human-written summaries");
console.log(" qmd ls [collection[/path]] - Inspect indexed files");
console.log("");
console.log("Remote LLM backend:");
console.log(" qmd remote set <urls> - Configure remote embed/rerank/generate endpoints");
console.log(" qmd remote status - Show remote config and check health");
console.log(" qmd remote clear - Clear remote config (revert to local)");
console.log("");
console.log("Maintenance:");
console.log(" qmd status - View index + collection health");
console.log(" qmd update [--pull] - Re-index collections (optionally git pull first)");
Expand Down Expand Up @@ -2592,6 +2650,7 @@ function showHelp(): void {
console.log("");
console.log("Global options:");
console.log(" --index <name> - Use a named index (default: index)");
console.log(" --local - Force local LLM mode (ignore remote config)");
console.log("");
console.log("Search options:");
console.log(" -n <num> - Max results (default 5, or 20 for --files/--json)");
Expand Down Expand Up @@ -2662,6 +2721,10 @@ if (isMain) {
process.exit(0);
}

if (cli.values.local) {
setForceLocalMode(true);
}

if (!cli.command || cli.values.help) {
showHelp();
process.exit(cli.values.help ? 0 : 1);
Expand Down Expand Up @@ -3132,14 +3195,100 @@ if (isMain) {
break;
}

case "remote": {
const subcommand = cli.args[0];
if (!subcommand) {
console.error("Usage: qmd remote <set|status|clear>");
console.error("");
console.error("Commands:");
console.error(" qmd remote set <embed-url> <rerank-url> <generate-url>");
console.error(" qmd remote status - Show current remote configuration");
console.error(" qmd remote clear - Clear remote config (use local mode)");
process.exit(1);
}

switch (subcommand) {
case "set": {
let embedUrl = cli.values["embed-url"] as string | undefined;
let rerankUrl = cli.values["rerank-url"] as string | undefined;
let generateUrl = cli.values["generate-url"] as string | undefined;
const generateModel = cli.values["generate-model"] as string | undefined;

if (cli.args.length >= 4) {
embedUrl = cli.args[1];
rerankUrl = cli.args[2];
generateUrl = cli.args[3];
} else if (cli.args.length === 2 && cli.args[1]) {
embedUrl = cli.args[1];
rerankUrl = cli.args[1];
generateUrl = cli.args[1];
}

if (!embedUrl && !rerankUrl && !generateUrl && !generateModel) {
console.error("Usage: qmd remote set <embed-url> <rerank-url> <generate-url>");
process.exit(1);
}

const existingConfig = loadRemoteConfig();
const newConfig = {
embedUrl: embedUrl ?? existingConfig.embedUrl,
rerankUrl: rerankUrl ?? existingConfig.rerankUrl,
generateUrl: generateUrl ?? existingConfig.generateUrl,
generateModel: generateModel ?? existingConfig.generateModel,
};

saveRemoteConfig(newConfig);
console.log("Remote configuration saved");
console.log(` Embed: ${newConfig.embedUrl || "(not set)"}`);
console.log(` Rerank: ${newConfig.rerankUrl || "(not set)"}`);
console.log(` Generate: ${newConfig.generateUrl || "(not set)"}`);
if (newConfig.generateModel) {
console.log(` Generate model: ${newConfig.generateModel}`);
}
break;
}
case "status": {
const config = loadRemoteConfig();
if (!config.embedUrl && !config.rerankUrl && !config.generateUrl) {
console.log("Remote mode: disabled (using local models)");
} else {
console.log("Remote Configuration\n");
console.log(` Embed: ${config.embedUrl || "(not set)"}`);
console.log(` Rerank: ${config.rerankUrl || "(not set)"}`);
console.log(` Generate: ${config.generateUrl || "(not set)"}`);
if (config.generateModel) {
console.log(` Generate model: ${config.generateModel}`);
}
console.log("\nChecking endpoint health...");
const remote = new RemoteLLM(config);
const health = await remote.checkHealth();
console.log(` Embed: ${health.embed ? "healthy" : "unreachable"}`);
console.log(` Rerank: ${health.rerank ? "healthy" : "unreachable"}`);
console.log(` Generate: ${health.generate ? "healthy" : "unreachable"}`);
}
break;
}
case "clear": {
clearRemoteConfig();
console.log("Remote configuration cleared");
console.log("Now using local models");
break;
}
default:
console.error(`Unknown remote subcommand: ${subcommand}`);
process.exit(1);
}
break;
}

default:
console.error(`Unknown command: ${cli.command}`);
console.error("Run 'qmd --help' for usage.");
process.exit(1);
}

if (cli.command !== "mcp") {
await disposeDefaultLlamaCpp();
await disposeAllLLM();
process.exit(0);
}

Expand Down
Loading