Omni-Search Audit & Roadmap (2026)

Omni-Search Audit & Roadmap (2026)

Status: Research document. This is a living audit. Its findings gesture toward an implementation plan that should be authored separately. Cross-reference: search-retrieval-ssot-2026.md for canonical pipeline contracts.

---

1. Executive Summary

Vox search stack is architecturally sound and significantly more capable than it appears at the GUI surface. The core vox-search crate implements a genuine hybrid multi-corpus planner (BM25 + vector embedding + RRF fusion) with 8 distinct backend legs and 6 SearchCorpus variants. The GUI exposes 7 user-facing scope chips and a federated client-side index for settings/commands.

However, major gaps exist between the backend power and what reaches users and agents:

1. The GUI SearchView does not surface scores, diagnostics, or backend-mix information meaningfully.
2. SearchCorpus::SymbolProximity exists but is never exercised by the GUI heuristic planner path.
3. There is no persistent AST/symbol index — repo search uses a slow WalkDir token-overlap scan (O(files x tokens)).
4. Chat search is a LIKE-only query bolted onto vox_search_query post-hoc, not a first-class corpus.
5. The Federated OmniSearch index (settings, surfaces, skills, commands, docs, policies, actions) lives entirely client-side and is rebuilt on every palette open with no fuzzy matching.
6. SearchCorpus has no variant for GitHistory, Telemetry, Plugins, Workflows, or AgentLogs.
7. The CommandPalette and SearchView share a useSearchController hook but have separate, slightly inconsistent filter logic.
8. No search telemetry pipeline: zero instrumentation of recall, precision, latency, or backend-mix ratios.
9. The repo_inventory_max_files: 20_000 cap means large repos silently truncate their file set.
10. No incremental indexing: every query re-walks the entire repo tree.

---

2. Current Search Architecture Map

2.1 Full Stack Layers

The search stack flows from top to bottom:

GUI Layer:

• CommandPalette (Cmd+K / Ctrl+K)
• SearchView (full search, sidebar)
• MemoryView (memory-scoped)

All GUI surfaces share useSearchController (200ms debounce) which calls vox_search_query via Tauri IPC. Federated surfaces (settings, commands) use useFederatedSearchIndex (client-side only).

Tauri command layer (vox-gui/src/commands/search.rs):

• vox_search_query command
• repo discovery (discover_repository_or_fallback)
• DB connect (optional, graceful degradation)
• heuristic_search_plan -> execute_search_plan
• chat_search_gui_messages (LIKE, post-hoc append)
• glob filter + kind filter + pagination

vox-search execution layer:

• SearchCorpus::Memory -> MemorySearchEngine (BM25+vector hybrid)
• SearchCorpus::KnowledgeGraph -> VoxDb::query_knowledge_nodes (FTS/LIKE)
• SearchCorpus::DocumentChunks -> VoxDb hybrid (FTS5 + embeddings)
• SearchCorpus::RepoInventory -> WalkDir token-overlap scan
• SearchCorpus::WebResearch -> SearXNG -> DDG -> Tavily
• SearchCorpus::SymbolProximity-> scan_symbol_proximity (rarely activated from GUI)
• [optional] Tantivy -> tantivy-lexical feature gate
• [optional] Qdrant -> qdrant-vector feature gate
• RRF fusion -> rrf_merge_line_lists (disabled by default)

vox-db / vox-db-types layer:

• SearchPlan, SearchCorpus, SearchBackend, heuristic_search_plan
• fuse_hybrid_results (BM25 + embedding score fusion)
• query_search_document_chunks_hybrid
• chat_search_gui_messages (LIKE only)

2.2 Corpus x Backend Matrix (Current State)

Corpus	Backend	GUI Accessible	Indexed	Incremental	Score Type
Memory	BM25 + optional vector	Yes	In-process	No (rebuild)	Hybrid f64
KnowledgeGraph	FTS/LIKE on knowledge_nodes	Yes	SQLite	No	Always 0.0 (BUG)
DocumentChunks	FTS5 + embedding hybrid	Yes	SQLite	Yes (ingest)	Hybrid f64
RepoInventory	WalkDir token-overlap	Yes	No index	No	Token count f32
WebResearch	SearXNG->DDG->Tavily	Yes (scope chip)	N/A	N/A	f64 from engine
SymbolProximity	scan_symbol_proximity	Partial (broken routing)	Partial	No	f64
Chats	LIKE on gui_messages	Yes (bolted on)	SQLite	Yes	Hardcoded 0.75
GitHistory	MISSING	No	No	No	—
PluginCatalog	Client-side federatedSearchIndex	Palette only	No	No	Exact/prefix
Settings	Client-side federatedSearchIndex	Palette + SearchView	No	No	Exact/prefix
Commands	Client-side filterCommandCatalogHits	Yes	No	No	Exact/contains
Telemetry	MISSING	No	No	No	—
AgentLogs	MISSING	No	No	No	—
Workflows	MISSING	No	No	No	—

2.3 GUI Search Surfaces

Surface	Entry Point	Backend Corpora	Notes
SearchView	Sidebar -> Search	All 7 scope chips	Full-featured; facets; path glob
CommandPalette	Cmd+K / Ctrl+K	Same backend + federated	Prefix modes /skills /commands /agents
MemoryView	Sidebar -> Memory	Memory-scoped	Direct memory search field
Chat preamble	Orchestrator (auto)	Memory + knowledge + chunks	Not visible in GUI
MCP vox_memory_search	Agent tools	Full retrieval bundle	Most powerful surface
MCP vox_knowledge_query	Agent tools	KnowledgeGraph only	Narrow
MCP vox_research_run	Agent tools	Web tier + CRAG	Research pipeline

---

3. Gaps & Bugs — Inventory

3.1 Critical Gaps

G1: No persistent AST/symbol index repo_path_search walks WalkDir for every query, scanning up to 20,000 files. On a repo with 121+ crates (as in Vox), this is O(50ms-500ms) per query. No tree-sitter parsing, no symbol extraction, no function/type/struct/trait graph. Impact: Code navigation queries return path matches, never symbol-level hits. SearchCorpus::SymbolProximity exists but uses scan_symbol_proximity which is a token proximity heuristic, not a real symbol graph.

G2: Chat corpus is second-class chat_search_gui_messages uses a LIKE %query% scan appended AFTER the main execution. No vector embedding, no BM25, hardcoded score of 0.75. Not part of SearchCorpus enum, not plannable, not controlled by heuristic plan. Chats are a primary user artifact but get the worst search quality.

G3: KnowledgeGraph score is always 0.0 execution.rs line 388: score: 0.0 for all knowledge nodes. These hits appear at the same RRF rank regardless of match quality. Makes knowledge graph results ungradable against memory/chunk results.

G4: Client-side federated index has no fuzzy matching scoreMatch in federatedSearchIndex.ts is exact/prefix only (100/80/50). Typo tolerance: zero. “setings” finds nothing in Settings. No trigram, Levenshtein, or phonetic matching. Palette is one of the most-used surfaces but has the crudest matching.

G5: SymbolProximity quality limited; excluded from default heuristic plan userScopeToBackend(‘code’) returns [‘repo’, ‘symbol’] (searchController.ts line 33-34). scope_to_corpus(‘symbol’) maps to SearchCorpus::SymbolProximity — so explicitly scoped code searches DO invoke it. However: (a) the default unscoped heuristic plan typically omits SymbolProximity, meaning users who do not select scope chips never see it; and (b) scan_symbol_proximity is a token-proximity heuristic (not a real symbol graph), so result quality is limited regardless of routing. Result: SymbolProximity is reachable via explicit code scope, but underutilized and low-quality in practice.

G6: No incremental repo indexing Every GUI search query restarts the full WalkDir scan. No inotify/FSEvents-based file watcher to maintain a persistent index. No content hash cache to skip unchanged files.

3.2 Important Gaps

G7: RRF fusion disabled by default policy.prefer_rrf_merge defaults to false. Without RRF, results are returned as separate per-corpus lists. Users see “memory”, then “chunk”, then “repo” — not interleaved by relevance. The GUI’s groupBySource groups by source, which hides cross-corpus ranking.

G8: Verification pass not wired in GUI path for scoped searches run_search_with_verification is called from vox_search_query only when scope_corpora is None. When user selects specific scope chips, execute_search_plan is called directly, skipping verification. Weak evidence is never detected in scoped searches.

G9: No search telemetry recordGamifyGuiEvent(‘search_query_executed’) fires but captures only query text. No latency tracking, no backend-mix logging, no click-through rate, no zero-results rate. Cannot improve ranking without usage signal.

G10: repo_inventory_max_files: 20_000 silent truncation Repos exceeding 20,000 files silently drop files. Vox itself has 150+ crates — this limit is precarious. No warning emitted to the UI when truncation occurs.

G11: Knowledge node scores unusable for RRF KnowledgeGraph hits enter RRF with score: 0.0. RRF is rank-based not score-based, so partially mitigated — but unified_hits sort still deprioritizes them.

G12: glob_match on frontend is naive and broken pathMatchesGlob in SearchView.tsx normalizes ** incorrectly:

• strips **/ prefix, replaces ** with ”, then * with ”
• ”**/*.rs” reduces to “.rs” (the extension only, with no wildcard)
• !normalized is false (string is non-empty), so the path.includes(‘.rs’) test runs
• Result: ”/*.rs” matches any path containing “.rs” anywhere, including paths like “foo.rss” or “bars/baz.rs” — it acts as a substring match on the extension, not a glob The correct behavior: ”/*.rs” should match only files ending in .rs at any depth. The backend glob_match in search.rs is correct (recursive DP); the frontend version is broken and should be replaced with a proper glob-to-regex conversion or deferred to the backend.

G13: chats scope causes superset results instead of chats-only (scope routing bug) userScopeToBackend(‘chats’) returns [‘chats’]. But scope_to_corpus(‘chats’) returns None — there is no SearchCorpus::Chats. When user selects only [‘chats’], scope_corpora maps to [] (empty, filtered to None), so the else branch runs run_search_with_verification with the full heuristic plan. Meanwhile, wants_chats is true (scope_tags contains ‘chats’), so chat results also append. Result: user asked for chats-only, but receives chat results PLUS all-corpora results — a strict superset. Chat messages are buried among unrelated memory/chunk/repo hits. This is more severe than “gets nothing”: the user cannot isolate chat history at all.

G14: Memory cache rebuild race cached_memory_engine caches the engine but if ctx.db is None, the engine is built from filesystem every call. Under concurrent GUI searches (debounced at 200ms), multiple rebuild calls can race.

G15: Web scraper disabled by default in GUI build The web-scrape feature gate is not enabled in vox-gui Cargo.toml dependency on vox-search. Web search results returned to GUI are engine snippets only (no full-page extraction). This makes web scope hits much lower quality than MCP/orchestrator-routed web searches.

3.3 Minor Issues

G16: Empty search query returns error instead of graceful empty state The Tauri command rejects empty strings; error message exposed in transport error logs.

G17: locator_for maps chunk to file but chunk IDs are not real file paths Chunk IDs from search_document_chunks are database row identifiers, not filesystem paths. Opening a “chunk” locator tries to spawn an editor with the chunk ID as a file path — fails silently.

G18: Score normalization inconsistent across corpora Memory BM25: unbounded float (TF-IDF based, can exceed 1.0). Chunk hybrid: normalized 0.0-1.0 via fusion. Repo token-overlap: count of matched tokens (integer-like f32). Web: engine-dependent (Tavily 0-1, SearXNG varies). Knowledge: always 0.0. Commands/Settings: hardcoded 0.85. scoreToPct in GUI divides by 1.0 then formats as percent — correct for normalized scores but garbage for BM25 values >1.0.

G19: No aria-live region for search result count updates Accessibility: screen readers do not announce when result count changes.

G20: CommandPalette does not share facet state with SearchView Selecting a scope in palette and pressing “View all results” should pre-populate SearchView with same scope — does not.

---

4. What Is Not Searched (Omni-Search Gaps)

The following content exists in the Vox codebase/runtime but is entirely unsearchable today:

Content Type	Location	Blocked By
Git commit history	.git/	No SearchCorpus::GitHistory
Git blame / authorship	git	No indexer
LSP diagnostic history	orchestrator logs	No corpus
Telemetry events (vox.script.*)	vox-telemetry	No corpus
Agent conversation logs	.gemini/antigravity/brain/	Not indexed
Workflow execution history	vox-journal	No corpus
Plugin manifest metadata	vox-plugin-catalog/catalog.toml	Client-side only
Skill YAML frontmatter	assets/skills/, .vox/skills/	Client-side only
ADR documents	docs/src/adr/	Not in memory corpus by default
Architecture docs	docs/src/architecture/	Not in memory corpus by default
Scientia findings	docs/src/architecture/	Not ingested
CHANGELOG / release notes	CHANGELOG.md	Not indexed
.env.example env var docs	.env.example	Not indexed
Cargo.toml feature flags	Cargo.toml per crate	No structured indexer
Contract YAML files	contracts/	Not indexed
MCP tool definitions	vox-orchestrator-mcp	Not indexed

---

5. Best Practices & 2026 Techniques

5.1 Indexing Architecture

The WalkDir-per-query approach is fundamentally unscalable. The 2026 standard:

• Tree-sitter-based symbol extraction at ingest time, building a symbol graph (function defs, struct defs, trait impls, macro uses). Vox has tree-sitter-vox already.
• Tantivy + content-hash cache for full-text: re-index only changed files by watching the filesystem (inotify on Linux, FSEvents on macOS, ReadDirectoryChangesW on Windows).
• Embedding vectors per symbol chunk (not per whole file) stored in SQLite (existing embeddings table).
• HNSW or flat ANN for small-to-medium corpora; Qdrant optional sidecar for production scale.

Key Rust tools:

• tantivy (already used) — mature Rust FTS, fast incremental indexing
• tree-sitter (already tree-sitter-vox) — structural parsing for symbol graph
• notify crate (inotify/FSEvents wrapper) — for filesystem watch events
• Content-addressable chunk store: hash file contents, only re-embed changed chunks

5.2 Hybrid Retrieval

RRF should be on by default in 2026. Research consistently shows hybrid BM25+vector+RRF outperforms any single leg by 10-30% NDCG. The prefer_rrf_merge flag should flip its default.

Score normalization: All corpus scores should be L2-normalized into [0, 1] before RRF and before display. Knowledge graph nodes need an actual match score (LIKE rank or BM25).

Query rewriting: The best_effort_verification_query is stopword-based. Query expansion with synonym sets or LLM query rewriting substantially improves recall on short queries.

5.3 Omni-Search UI Patterns (2026)

Linear, Notion, and Raycast have converged on:

• Sub-50ms first result from local indexes (federated client-side index or pre-built local SQLite)
• Progressive loading: show local/cached results instantly, then stream in backend results
• Intent-based routing: parse @memory, @code, #tag, >command prefixes to route to specific corpus
• Keyboard-centric: full arrow/enter/escape with no mouse required
• Result preview panel: hover or right-side panel showing full content before clicking
• Fuzzy matching with Levenshtein: tolerance for 1-2 typos in short queries
• Recent/pinned results: surfacing frequently-accessed items even without a query
• Search as navigation: pressing Enter opens the file in the editor, not just copies path

5.4 Agent-Facing Search (2026)

For AI agents working in a codebase, the most useful search capabilities are:

• Symbol definition lookup: “where is SearchPolicy defined?” -> immediate answer
• Call graph traversal: “what calls execute_search_plan?” -> callers list
• Semantic code search: “find functions that do hybrid fusion” -> embedding-based
• Change-aware search: “what changed in this crate recently?” -> git-blame-aware
• Dependency graph: “what crates depend on vox-search?” -> Cargo.lock parse
• Diagnostic search: “find all files with recent compiler warnings” -> telemetry-backed

The MCP surface (vox_memory_search, vox_knowledge_query) is strong but missing symbol-level and call-graph queries.

5.5 Performance Benchmarks (2026 Targets)

Query Type	Current (estimated)	Target
Memory BM25 (300 docs)	~10ms	<5ms
Repo path walk (20k files)	50-500ms	<20ms (cached index)
Chunk hybrid (FTS5 + embedding)	20-100ms	<30ms
Web search (SearXNG)	500-2000ms	500ms (unchanged)
Federated palette (client)	<5ms	<5ms
Full omni-search (all corpora)	~200-800ms	<100ms local, progressive

---

6. Improvements Inventory (Prioritized)

Tier 1 — High Impact, Moderate Effort (Fix Now)

ID	Improvement	Est. Effort
I-01	Fix chats scope mapping bug (G13) — chats scope fans out to all corpora	S
I-02	Fix glob_match frontend broken for ** patterns (G12)	S
I-03	Return real KnowledgeGraph scores from query_knowledge_nodes (G3)	M
I-04	Enable prefer_rrf_merge by default (G7)	XS
I-05	Add fuzzy matching (trigram/Levenshtein) to federatedSearchIndex (G4)	M
I-06	Fix locator_for chunk->file mapping — chunk IDs are not paths (G17)	M
I-07	Add aria-live to search result count (G19)	XS
I-08	Wire SymbolProximity into GUI code scope reliably (G5)	M
I-09	Emit warning when repo files truncated at max (G10)	S
I-10	Normalize BM25 scores to [0,1] before display (G18)	M

Tier 2 — High Impact, Higher Effort

ID	Improvement	Est. Effort
I-11	Persistent repo file index with content-hash cache	L
I-12	Filesystem watcher for incremental index updates	L
I-13	Chat corpus as first-class SearchCorpus::Chats	L
I-14	Tree-sitter symbol extraction index	XL
I-15	Search telemetry: latency + backend-mix + zero-results (G9)	M
I-16	Progressive/streaming search results in UI	L
I-17	Result preview panel on hover	M
I-18	Enable web-scrape feature in GUI build (G15)	S
I-19	Query intent prefix routing (@memory, #tag, >cmd)	M
I-20	Open in editor for repo/chunk hits (not just copy path)	M

Tier 3 — New Corpora

ID	Improvement	Est. Effort
I-21	SearchCorpus::GitHistory — commit/blame search	XL
I-22	SearchCorpus::TelemetryEvents — query past events	L
I-23	SearchCorpus::AgentLogs — conversation history search	L
I-24	Ingest docs/src/ into DocumentChunks at startup	M
I-25	Ingest contracts/ YAML into knowledge graph	M
I-26	Index Cargo.toml feature flags as structured metadata	M
I-27	MCP tool for symbol definition lookup	L
I-28	MCP tool for call graph / callers query	XL

---

7. SSOT Refinements Required

The existing search-retrieval-ssot-2026.md needs the following additions:

1. Add SearchCorpus::Chats — or explicitly document that chats are a post-hoc corpus handled outside the standard planner, with justification.
2. Document SymbolProximity routing gap — the GUI code scope maps to symbol but this does not reliably invoke the corpus.
3. Add GUI surface table — SSOT section 6 does not document the vox_search_query Tauri command or the federated index.
4. Specify score normalization policy — what range should hit scores be in? The SSOT is silent.
5. Document feature gate matrix — tantivy-lexical, qdrant-vector, tavily, web-scrape feature gates and which surfaces activate which.
6. Add federated index contract — contracts/gui/omnisearch-index.v1.yaml exists but is not referenced in the SSOT.

---

8. Code Review Findings (Pre-Submit)

Issues identified during this audit that should be caught before any search-related changes are merged:

BUG: chats scope silently fans out to all corpora (search.rs:295)

When scope = [“chats”]:

• scope_corpora maps “chats” through scope_to_corpus -> None -> filtered out -> scope_corpora = Some([]) -> treated as None
• The else branch calls run_search_with_verification with the FULL heuristic plan
• wants_chats is true (scope_tags contains “chats”), so chat results append (lines 295-318)
• User asked for chats-only, but receives chat results PLUS all-corpora results (superset)
• Chat messages are indistinguishable from memory/chunk/repo hits in the result list

Fix: Add SearchCorpus::Chats to make chats a plannable corpus, or short-circuit: detect scope_tags == [“chats”] before calling run_search_with_verification and route only to chat_search_gui_messages.

BUG: Frontend glob pathMatchesGlob is broken for ** (SearchView.tsx:45-52)

const normalized = pattern.replace(/^**//, ”).replace(/**/g, ”).replace(/*/g, ”);

“/.rs” -> “.rs” -> “.rs” (after star removal, the leading * is consumed but the dot and extension remain) -> normalized is “.rs” (truthy) -> path.includes(‘.rs’) -> matches any path containing “.rs” anywhere. This means ”/*.rs” acts as a substring match for “.rs”, catching file paths like “bars/baz.rs” AND paths containing “.rs” in directory names or other extensions — not a proper recursive glob. The correct behavior requires full glob-to-regex conversion.

Fix: Move glob matching to the backend glob_match in search.rs (which is correct via recursive DP), passing the path_glob parameter through the existing vox_search_query scope filter, or implement proper glob-to-regex conversion in TypeScript (e.g., micromatch or a simple recursive DP port of the backend implementation).

BUG: KnowledgeGraph score always 0.0 (execution.rs:388)

score: 0.0,

Knowledge node hits are inserted into unified_hits with score 0.0, placing them at the bottom of any score-sorted result list regardless of query match quality. The query_knowledge_nodes DB function already does FTS/LIKE ranking — the rank should be propagated.

Fix: Extend VoxDb::query_knowledge_nodes to return a score or rank; propagate it in execution.rs.

LATENT BUG: Memory engine race on concurrent searches (memory_cache.rs)

The cached_memory_engine function uses a process-global cache. When two GUI searches fire within the 200ms debounce window, both may construct separate MemorySearchEngine instances from disk.

Fix: Ensure the cache uses Arc<RwLock<>> with proper async locking; return the same instance for concurrent callers.

---

9. 2026 Roadmap Gestures

This section sketches the direction of a full implementation plan (to be authored separately as omni-search-implementation-plan-2026.md).

Phase A: Fix & Consolidate (Weeks 1-2)

Fix all Tier 1 bugs (G12, G13, G3) and enable RRF by default. Pure fixes with no new infrastructure. Unlocks correct behavior from the existing stack.

Phase B: Telemetry + Score Normalization (Weeks 2-3)

Instrument every search call with backend-mix, latency, result count. Normalize scores across corpora. Add aria-live accessibility. Enables data-driven ranking improvements.

Phase C: Persistent Repo Index (Weeks 3-6)

Build a content-hash-cached Tantivy index for repo files with tree-sitter symbol extraction. Add filesystem watcher for incremental updates. Graduate repo_inventory_max_files limit to a soft warning with UI indicator.

Phase D: Chat as First-Class Corpus (Weeks 4-5)

Add SearchCorpus::Chats to the planner. Embed chat messages at ingest time. Add hybrid BM25+vector search for chats. Fix the scope mapping bug.

Phase E: Omni-Search UX Elevation (Weeks 5-8)

Progressive streaming results. Result preview panel. Intent prefix routing. Fuzzy matching in federated index. “Open in editor” for file hits. Share facet state between palette and SearchView.

Phase F: Agent-Facing Corpus Expansion (Weeks 8-12)

Git history corpus. Symbol definition MCP tool. Call-graph traversal. Automatic ingestion of docs/src/ and contracts/ into knowledge graph.

---

10. Related Documents

• search-retrieval-ssot-2026.md — canonical pipeline contracts (update after findings here)
• data-storage-ssot-2026.md — DB schema including search_documents, embeddings
• vox-gui-capability-audit-2026.md — GUI surface audit
• vox-gui-surface-map-2026-06-14.md — surface map
• deep-research-prior-art-and-vox-roadmap-2026.md — web research pipeline prior art

---

Appendix A: Feature Gate Matrix

Feature Gate	Default	Enables	GUI Build	MCP/CLI Build
tantivy-lexical	Off	Tantivy doc mirror index	No	Optional
qdrant-vector	On	Qdrant ANN sidecar	If URL set	Yes
tavily	On	Tavily web tier	If API key set	Yes
web-scrape	Off	Full-page HTML extraction	MISSING	Optional

Note: web-scrape being off in the GUI build means web search results in SearchView are snippets only (30-100 chars) while the same query via MCP returns full extracted markdown. This is a significant quality disparity.

---

Appendix B: Env / Secret Knobs Affecting GUI Search

Secret / Env	Default	Effect
VOX_SEARCH_PREFER_RRF	false	Cross-corpus fusion — SHOULD be true
VOX_SEARCH_MEMORY_VECTOR_WEIGHT	0.55	Memory BM25/vector blend
VOX_SEARCH_CHUNK_VECTOR_WEIGHT	0.60	Chunk FTS/embedding blend
VOX_SEARCH_BM25_K1	1.2	Memory BM25 saturation
VOX_SEARCH_BM25_B	0.75	Memory BM25 length norm
VOX_SEARCH_RRF_K	60.0	RRF smoothing constant
VOX_SEARCH_REPO_MAX_FILES	20000	File scan cap (often too low)
VOX_SEARCH_QDRANT_URL	unset	Qdrant sidecar ANN
VOX_SEARCH_TANTIVY_ROOT	unset	Tantivy doc mirror
VOX_SEARCH_SEARXNG_URL	unset	SearXNG instance
VOX_SEARCH_TAVILY_ENABLED	false	Tavily web tier
VOX_SEARCH_PERSIST_WEB_HITS_DISABLED	unset	Web hit caching to DB