@gmickel/gno 1.34.0 → 1.34.2

package/README.md

@@ -117,7 +117,7 @@ gno daemon --detach  # headless indexing + resident MCP gateway
 
 <!-- public-truth:current-version -->
 
-> Current release: **v1.34.0** — see [CHANGELOG.md](./CHANGELOG.md)
+> Current release: **v1.34.1** — see [CHANGELOG.md](./CHANGELOG.md)
 
 <!-- /public-truth -->
 

package/browser-extension/artifacts/gno-browser-clipper-v1.34.2.zip.sha256

@@ -0,0 +1 @@
+ee0e3d4032a41a097c85fba552a9ab449906f258e14844ad5e3b5f99001716e0  gno-browser-clipper-v1.34.2.zip

package/browser-extension/dist/manifest.json

@@ -21,5 +21,5 @@
   "content_security_policy": {
     "extension_pages": "script-src 'self'; object-src 'none'; connect-src http://127.0.0.1:*"
   },
-  "version": "1.34.0"
+  "version": "1.34.2"
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gmickel/gno",
-  "version": "1.34.0",
+  "version": "1.34.2",
   "description": "Local semantic search for your documents. Index Markdown, PDF, and Office files with hybrid BM25 + vector search.",
   "keywords": [
     "embeddings",

package/spec/cli.md

@@ -1922,6 +1922,8 @@ gno models pull [--all|--embed|--rerank|--gen] [--force]
 **Behavior:**
 
 - Skips models that are already cached (checksum match) unless `--force` is used
+- Skips HTTP(S) rerank endpoints as external services; they are called directly
+  and are never downloaded or cached, including with `--force`
 - Default (no flags): pulls all models
 
 **Exit Codes:**

package/src/cli/commands/models/pull.ts

@@ -10,6 +10,7 @@ import type { DownloadProgress, ModelType } from "../../../llm/types";
 import { getModelsCachePath } from "../../../app/constants";
 import { loadConfig } from "../../../config";
 import { ModelCache } from "../../../llm/cache";
+import { isHttpRerankUri } from "../../../llm/httpRerank";
 import { getActivePreset } from "../../../llm/registry";
 
 // ─────────────────────────────────────────────────────────────────────────────
@@ -46,6 +47,7 @@ export interface ModelPullResult {
   error?: string;
   path?: string;
   skipped?: boolean;
+  skipReason?: "cached" | "external";
 }
 
 export interface ModelsPullResult {
@@ -117,6 +119,20 @@ export async function modelsPull(
     const uri =
       type === "expand" ? (preset.expand ?? preset.gen) : preset[type];
 
+    // HTTP rerankers are services, not model artifacts. They are loaded
+    // directly by LlmAdapter and must never enter the local model cache path.
+    if (type === "rerank" && isHttpRerankUri(uri)) {
+      results.push({
+        type,
+        uri,
+        ok: true,
+        skipped: true,
+        skipReason: "external",
+      });
+      skipped += 1;
+      continue;
+    }
+
     // Check if already cached (skip unless --force)
     if (!options.force) {
       const isCached = await cache.isCached(uri);
@@ -128,6 +144,7 @@ export async function modelsPull(
           ok: true,
           path: path ?? undefined,
           skipped: true,
+          skipReason: "cached",
         });
         skipped += 1;
         continue;
@@ -176,13 +193,17 @@ export async function modelsPull(
  */
 export function formatModelsPull(result: ModelsPullResult): string {
   const lines: string[] = [];
+  let externalSkipped = 0;
   const label = (type: ModelType) =>
     type === "gen" ? "answer" : type === "expand" ? "expand" : type;
 
   for (const r of result.results) {
     if (r.ok) {
       if (r.skipped) {
-        lines.push(`${label(r.type)}: skipped (already cached)`);
+        if (r.skipReason === "external") externalSkipped += 1;
+        const reason =
+          r.skipReason === "external" ? "external endpoint" : "already cached";
+        lines.push(`${label(r.type)}: skipped (${reason})`);
       } else {
         lines.push(`${label(r.type)}: downloaded`);
       }
@@ -196,10 +217,18 @@ export function formatModelsPull(result: ModelsPullResult): string {
     lines.push(`${result.failed} model(s) failed to download.`);
   } else if (result.skipped === result.results.length) {
     lines.push("");
-    lines.push("All models already cached. Use --force to re-download.");
+    lines.push(
+      externalSkipped > 0
+        ? "No model downloads needed."
+        : "All models already cached. Use --force to re-download."
+    );
   } else {
     lines.push("");
-    lines.push("All models downloaded successfully.");
+    lines.push(
+      externalSkipped > 0
+        ? "All downloadable models downloaded successfully."
+        : "All models downloaded successfully."
+    );
   }
 
   return lines.join("\n");

package/src/llm/nodeLlamaCpp/generation.ts

@@ -60,6 +60,29 @@ const DEFAULT_TEMPERATURE = 0;
 const DEFAULT_SEED = 42;
 const DEFAULT_MAX_TOKENS = 256;
 
+// Context sizing: without an explicit contextSize, node-llama-cpp defaults to
+// "auto", which grows the KV cache to fill available VRAM up to the model's
+// trained context length (OOM risk on small GPUs — see issue #189). Instead,
+// size the context to what the call actually needs: prompt tokens + output
+// budget + margin for chat-template wrapping and special tokens.
+const GEN_CONTEXT_MARGIN_TOKENS = 512;
+const GEN_CONTEXT_MIN_TOKENS = 1024;
+
+export const resolveGenContextSize = (input: {
+  promptTokenCount: number;
+  maxTokens: number;
+  trainContextSize?: number;
+}): number => {
+  const needed = Math.max(
+    GEN_CONTEXT_MIN_TOKENS,
+    input.promptTokenCount + input.maxTokens + GEN_CONTEXT_MARGIN_TOKENS
+  );
+  if (input.trainContextSize && input.trainContextSize > 0) {
+    return Math.min(needed, input.trainContextSize);
+  }
+  return needed;
+};
+
 // ─────────────────────────────────────────────────────────────────────────────
 // Implementation
 // ─────────────────────────────────────────────────────────────────────────────
@@ -97,9 +120,14 @@ export class NodeLlamaCppGeneration implements GenerationPort {
             await this.manager.getLlama()
           ).createGrammarForJsonSchema(params.jsonSchema as JsonGrammarSchema)
         : undefined;
-      context = await llamaModel.createContext(
-        params?.contextSize ? { contextSize: params.contextSize } : undefined
-      );
+      const contextSize =
+        params?.contextSize ??
+        resolveGenContextSize({
+          promptTokenCount: llamaModel.tokenize(prompt).length,
+          maxTokens: params?.maxTokens ?? DEFAULT_MAX_TOKENS,
+          trainContextSize: llamaModel.trainContextSize,
+        });
+      context = await llamaModel.createContext({ contextSize });
       // Import LlamaChatSession dynamically
       const { LlamaChatSession } = await import("node-llama-cpp");
       const session = new LlamaChatSession({

package/src/store/sqlite/adapter.ts

@@ -128,11 +128,7 @@ import {
   classifyResolvedGraphEdge,
   mergeGraphEdgeAudit,
 } from "../../core/graph-edge-confidence";
-import {
-  buildWikiBestMatchSubquery,
-  buildWikiBestRankMatchCountSubquery,
-  buildWikiBestRankSubquery,
-} from "../../core/graph-resolver";
+import { buildWikiBestMatchSubquery } from "../../core/graph-resolver";
 import { buildContentPrefilterNeedles } from "../../core/link-relevance";
 import { normalizeWikiName, stripWikiMdExt } from "../../core/links";
 import {
@@ -179,6 +175,7 @@ import {
   getLatestFileRefactorReceiptByPlanDigest as getStoredLatestFileRefactorReceiptByPlanDigest,
 } from "./file-refactor-journal-store";
 import { loadFts5Snowball } from "./fts5-snowball";
+import { resolveGraphLinkTargets } from "./graph-link-resolver";
 import { queryGraphNeighborsForSeeds } from "./graph-neighbors";
 import {
   appendExportManifest as appendStoredTraceExportManifest,
@@ -4757,9 +4754,13 @@ export class SqliteAdapter implements StorePort, SqliteDbProvider {
         match_count: number | null;
       }
 
-      interface UnresolvedByTypeRow {
+      interface GraphLinkResolutionRow {
+        source_id: number;
+        source_docid: string;
+        source_collection: string;
+        target_ref_norm: string;
+        target_collection: string | null;
         link_type: "wiki" | "markdown";
-        unresolved: number;
       }
 
       interface NodeMetaRow {
@@ -4771,104 +4772,68 @@ export class SqliteAdapter implements StorePort, SqliteDbProvider {
         rel_path: string;
       }
 
-      const edgeParams: string[] = [];
-      let edgeCollectionClause = "";
+      const linkParams: string[] = [];
+      let sourceCollectionClause = "";
       if (collection) {
-        edgeCollectionClause = "AND src.collection = ? AND tgt.collection = ?";
-        edgeParams.push(collection, collection);
+        sourceCollectionClause = "AND src.collection = ?";
+        linkParams.push(collection);
       }
 
-      const resolvedEdgeQuery = `
+      const graphLinkRows = db
+        .query<GraphLinkResolutionRow, string[]>(
+          `
         SELECT
           src.id as source_id,
           src.docid as source_docid,
-          tgt.id as target_id,
-          tgt.docid as target_docid,
-          dl.link_type,
-          CASE dl.link_type
-            WHEN 'wiki' THEN (${buildWikiBestRankSubquery(
-              "COALESCE(dl.target_collection, src.collection)",
-              "dl.target_ref_norm"
-            )})
-            WHEN 'markdown' THEN 5
-          END as match_rank,
-          CASE dl.link_type
-            WHEN 'wiki' THEN (${buildWikiBestRankMatchCountSubquery(
-              "COALESCE(dl.target_collection, src.collection)",
-              "dl.target_ref_norm"
-            )})
-            WHEN 'markdown' THEN 1
-          END as match_count
+          src.collection as source_collection,
+          dl.target_ref_norm,
+          dl.target_collection,
+          dl.link_type
         FROM documents src
         JOIN doc_links dl ON dl.source_doc_id = src.id
-        JOIN documents tgt ON tgt.id = CASE dl.link_type
-          WHEN 'wiki' THEN (${buildWikiBestMatchSubquery(
-            "COALESCE(dl.target_collection, src.collection)",
-            "dl.target_ref_norm"
-          )})
-          WHEN 'markdown' THEN (
-            SELECT md.id FROM documents md
-            WHERE md.active = 1
-              AND md.collection = COALESCE(dl.target_collection, src.collection)
-              AND md.rel_path = dl.target_ref_norm
-            ORDER BY md.id LIMIT 1
-          )
-        END
-        WHERE src.active = 1 AND tgt.active = 1
-          ${edgeCollectionClause}
-        ORDER BY src.id ASC, tgt.id ASC, dl.link_type ASC
-      `;
-
-      const resolvedEdgeRows = db
-        .query<ResolvedEdgeRow, string[]>(resolvedEdgeQuery)
-        .all(...edgeParams);
-
-      const unresolvedParams: string[] = [];
-      let unresolvedCollectionClause = "";
-      if (collection) {
-        unresolvedCollectionClause = "AND src.collection = ?";
-        unresolvedParams.push(collection);
-      }
-      const unresolvedQuery = `
-        SELECT
-          link_type,
-          COUNT(*) as unresolved
-        FROM (
-          SELECT
-            dl.link_type,
-            CASE dl.link_type
-              WHEN 'wiki' THEN (
-                ${buildWikiBestMatchSubquery(
-                  "COALESCE(dl.target_collection, src.collection)",
-                  "dl.target_ref_norm"
-                )}
-              )
-              WHEN 'markdown' THEN (
-                SELECT t.id FROM documents t
-                WHERE t.active = 1
-                  AND t.collection = COALESCE(dl.target_collection, src.collection)
-                  AND t.rel_path = dl.target_ref_norm
-                ORDER BY t.id LIMIT 1
-              )
-            END as target_id
-          FROM documents src
-          JOIN doc_links dl ON dl.source_doc_id = src.id
-          WHERE src.active = 1
-            ${unresolvedCollectionClause}
+        WHERE src.active = 1
+          ${sourceCollectionClause}
+        ORDER BY src.id ASC, dl.id ASC
+      `
         )
-        WHERE target_id IS NULL
-        GROUP BY link_type
-      `;
-      const unresolvedRows = db
-        .query<UnresolvedByTypeRow, string[]>(unresolvedQuery)
-        .all(...unresolvedParams);
+        .all(...linkParams);
+      const resolutions = resolveGraphLinkTargets(
+        db,
+        graphLinkRows.map((row) => ({
+          targetRefNorm: row.target_ref_norm,
+          targetCollection: row.target_collection ?? row.source_collection,
+          linkType: row.link_type,
+        }))
+      );
+      const resolvedEdgeRows: ResolvedEdgeRow[] = [];
       const unresolvedByType: Record<"wiki" | "markdown", number> = {
         wiki: 0,
         markdown: 0,
       };
-      for (const row of unresolvedRows) {
-        unresolvedByType[row.link_type] = row.unresolved;
+      for (const [index, row] of graphLinkRows.entries()) {
+        const resolution = resolutions[index];
+        if (!resolution) {
+          unresolvedByType[row.link_type] += 1;
+          continue;
+        }
+        const targetCollection = row.target_collection ?? row.source_collection;
+        if (collection && targetCollection !== collection) continue;
+        resolvedEdgeRows.push({
+          source_id: row.source_id,
+          source_docid: row.source_docid,
+          target_id: resolution.targetId,
+          target_docid: resolution.targetDocid,
+          link_type: row.link_type,
+          match_rank: resolution.matchRank,
+          match_count: resolution.matchCount,
+        });
       }
+      resolvedEdgeRows.sort(
+        (left, right) =>
+          left.source_id - right.source_id ||
+          left.target_id - right.target_id ||
+          left.link_type.localeCompare(right.link_type)
+      );
       const totalEdgesUnresolved =
         unresolvedByType.wiki + unresolvedByType.markdown;
 

package/browser-extension/artifacts/gno-browser-clipper-v1.34.0.zip.sha256

@@ -1 +0,0 @@
-0acc125d848721bdced540fe49412ef8c83df5d4ff3ba086076c46ed9d441998  gno-browser-clipper-v1.34.0.zip