Vox Deep Research Capabilities — Full Audit & 2026 Roadmap

Vox Deep Research Capabilities — Full Audit & 2026 Roadmap

Date: 2026-06-17
Scope: Search stack, research pipeline, novelty detection, OpenRouter free tier, CRAG / SCIENTIA integration — current state vs. 2026 SOTA, with gap analysis and phased implementation plan.

2026-08-01 update: This audit’s priority ordering is superseded by deep-research-synthesis-and-priorities-2026-08-01.md. Each of the 9 gaps below (G1–G9) was independently re-verified against current code in deep-research-verification-2026-08-01.md — most of G1/G4/G5/G6 shipped within 48 hours of this doc, then stalled; see that doc for exact status. This audit’s technical content otherwise remains a valid baseline.

---

1. Executive Summary

Vox has a substantially complete deep-research scaffold today. The pipeline exists end-to-end: query decomposition → multi-source retrieval (local hybrid + web dispatcher) → CRAG multi-hop loop → claim extraction → LLM-as-judge quality scoring → SCIENTIA finding-candidate promotion. The critical delta between Vox’s current state and production-grade 2026 deep research is a set of well-scoped gaps in:

1. LLM-driven reflection — CRAG query expansion is heuristic regex-only, not LLM-generated
2. Novelty / deduplication scoring — no per-hit information-gain gating; URL dedup is basic
3. Cross-encoder reranking — the hybrid fusion uses RRF but there is no dedicated reranking model
4. Confidence gate — score_with_config is a Phase 0a citation-count stub; multi-signal fusion not yet wired
5. Free-tier OpenRouter model routing — openrouter/free virtual model exists but research cascade defaults to openrouter/auto (paid-tier)
6. Synthesis context budget — max_tokens for Synthesis = 1,800 is too small for long-form deep research reports
7. Async / durable research jobs — session tracking exists but durable execution is not production-complete

Filling these gaps in order of leverage will complete a production-grade, cost-aware deep research loop running on the existing Vox infrastructure.

---

2. State-of-the-Art Benchmark (2026)

2.1 Agentic Deep Research Loop (PRAR Cycle)

2026 SOTA deep research agents implement a Perceive → Reason → Act → Reflect (PRAR) iterative loop:

Phase	What it does
Plan	Decompose research goal into 3–12 precise sub-queries using an LLM with JSON schema output
Retrieve	Hybrid retrieval: dense ANN vector search + sparse BM25, fused via RRF (k=60); live web search via policy-gated backends
Rerank	Cross-encoder pass on top-200 candidates → keep top-20 for synthesis
Reflect	LLM evaluates: “Is this evidence sufficient? What is missing? What contradicts?”
Re-plan	If gaps found, generate novel follow-up queries and loop; CRAG stop condition: quality ≥ target
Verify	NLI-class model checks each extracted claim against retrieved evidence (Support / Contradict / Contested / Unverified)
Synthesize	Cited answer with evidence spans
Judge	LLM-as-judge quality score (factual accuracy + citation density + coverage)
Promote	If quality ≥ threshold, promote to SCIENTIA finding candidate

2.2 2026 Retrieval Infrastructure Best Practices

Component	2026 Best Practice	Vox Current State
Sparse index	BM25 (in-process or Tantivy)	✅ BM25 in-process (memory_hybrid.rs)
Dense index	Embedding vectors (MiniLM / nomic-embed)	✅ Optional Qdrant ANN (vector_qdrant.rs)
Fusion	Reciprocal Rank Fusion (RRF k=60)	✅ rrf.rs — prefer_rrf_merge flag
Reranking	Cross-encoder on top-200 candidates	❌ Not implemented
Web search tier	SearXNG → DDG → Tavily → scrape	✅ web_dispatcher.rs — all 4 tiers
Deep research API	Tavily /research endpoint	✅ tavily_research.rs (gated by VOX_TAVILY_RESEARCH)
Multi-hop	Quality-gated CRAG expansion	✅ crag.rs + research.rs
Novelty filtering	Per-hit information-gain scoring	❌ Only URL-level dedup
Semantic caching	Intent-based result caching	⚠️ pipeline_cache.rs exists (TTL cache only)
Observability	Per-hop span tracing	⚠️ Telemetry events exist; no per-hop span

2.3 OpenRouter Free Tier (June 2026)

OpenRouter provides 20+ permanently-free models via the :free suffix or the openrouter/free router:

Model slug	Context	Strengths	RPD limit
google/gemma-3-27b-it:free	131K	Instruction following	~200
deepseek/deepseek-r1:free	164K	Reasoning	~50
meta-llama/llama-3.3-70b-instruct:free	131K	General	~200
microsoft/phi-4-reasoning:free	16K	Chain-of-thought	~200
openrouter/free (router)	varies	Auto-selects	aggregated

Key API facts: OpenAI-compatible; base_url = https://openrouter.ai/api/v1; free-tier users get ~50–200 RPD; $10 credit purchase unlocks ~1,000 RPD. Reasoning models support reasoning parameter.

---

3. Vox Current Architecture — Full Audit

3.1 Layer Map

┌─────────────────────────────────────────────────────────────────┐
│  GUI / CLI / MCP surfaces                                       │
│  (vox research run, vox_research_run MCP tool)                 │
├─────────────────────────────────────────────────────────────────┤
│  vox-research-shim  (orchestrator/pipeline.rs)                  │
│  ├── planner.rs  ← LLM decompose (cascade: local → OpenRouter) │
│  ├── web_gather.rs ← local hybrid + web via vox-search          │
│  ├── claims.rs ← claim extraction (LLM cascade)                 │
│  ├── verifier.rs ← NLI-style verdict (LLM cascade)             │
│  ├── stages.rs ← judge_quality(), synthesize_answer()          │
│  ├── gate.rs ← confidence score → routing tier                  │
│  ├── discovery_bridge.rs ← SCIENTIA finding candidate promotion │
│  └── pipeline_cache.rs ← TTL-based short-circuit cache         │
├─────────────────────────────────────────────────────────────────┤
│  vox-search  (shared by MCP, orchestrator, CLI, A2A)           │
│  ├── execution.rs ← SearchExecution (all 8 corpora)            │
│  ├── memory_hybrid.rs ← BM25 + optional vector fusion          │
│  ├── rrf.rs ← Reciprocal Rank Fusion                           │
│  ├── crag.rs ← CragRouter (query expansion, stop condition)    │
│  ├── research.rs ← run_multi_hop_web_research()                │
│  ├── web_dispatcher.rs ← SearXNG → DDG → Tavily → scrape      │
│  ├── tavily_research.rs ← /research deep-research endpoint     │
│  ├── bundle.rs ← run_search_with_verification()                │
│  ├── policy.rs ← SearchPolicy (all tunables)                   │
│  └── evaluation.rs ← SearchEvalReport                          │
├─────────────────────────────────────────────────────────────────┤
│  vox-actor-runtime  (LLM facade)                               │
│  ├── llm/cascade.rs ← cascade_for_research_stage()            │
│  │   → local Ollama/Populi → OpenRouter (via key)             │
│  └── model_resolution.rs ← RouteResolutionInput               │
├─────────────────────────────────────────────────────────────────┤
│  vox-research-shim/selection/  (model selection)               │
│  ├── free_tier.rs ← FreeTierRouter                             │
│  ├── virtual_models.rs ← openrouter/auto, openrouter/free      │
│  └── scorer.rs ← ModelScorer                                   │
├─────────────────────────────────────────────────────────────────┤
│  vox-scientia  (knowledge platform)                            │
│  ├── claim_extractor/ ← claim parsing from synthesized text    │
│  ├── critic_gate/ ← quality gate for publications             │
│  ├── review_flow.rs ← human-gated discovery review            │
│  └── evidence_assist.rs ← LLM evidence/conclusion suggestions  │
└─────────────────────────────────────────────────────────────────┘

3.2 Research Pipeline Stages (production state)

Stage	Status	Code Location
Session creation	✅ Shipped	pipeline.rs:100-111
Query decomposition (LLM)	✅ Shipped	planner.rs:decompose_query_with_config
Local hybrid retrieval	✅ Shipped	web_gather.rs:gather_local_hits_for_plan
Web retrieval (SearXNG/DDG/Tavily)	✅ Shipped	web_gather.rs:gather_web_hits_for_plan
URL-level deduplication	✅ Shipped	pipeline.rs:dedupe_hits_by_url
Retrieval diagnostics	✅ Shipped	pipeline.rs:217-259
Claim extraction	✅ Shipped	claims.rs:extract_claims_with_model
Confidence gate (routing tier)	⚠️ Phase 0a stub	gate.rs:score_with_config — citation count only
Claim verification (NLI)	✅ Shipped	verifier.rs:verify_claims_with_config — only on DeepResearch tier
Answer synthesis	✅ Shipped	stages.rs:synthesize_answer_with_llm
LLM-as-judge quality scoring	✅ Shipped	stages.rs:judge_quality
Self-verification (CoVE)	✅ Shipped	stages.rs:run_self_verification
Citation auditing	✅ Shipped	pipeline.rs:citation_audit
SCIENTIA promotion	✅ Shipped	discovery_bridge.rs
Report persistence	✅ Shipped	pipeline.rs:persist_enabled
MENS training pair generation	✅ Shipped	pipeline.rs:training_pair_min_confidence

3.3 Web Search Dispatcher Stack

WebSearchDispatcher::search(query, policy)
  ├── Tier 1: [Reserved/internal]
  ├── Tier 2: SearXNG (policy.searxng_url)  ← free, self-hosted
  ├── Tier 3: DuckDuckGo (policy.duckduckgo_fallback_enabled)  ← free
  └── Tier 4: Tavily search (policy.tavily_enabled, VOX_TAVILY_API_KEY)  ← paid

  Post-retrieval:
  ├── rank_and_dedupe_results()  ← source_authority_score (gov/edu/arxiv/github boost)
  ├── tavily_extract::uplift_low_quality_snippets()  ← content enrichment
  └── scraper::fetch_and_extract()  ← optional full-page extraction (web-scrape feature)

Multi-hop:
  run_multi_hop_web_research()  ← research.rs
    ├── hop loop: CragRouter::expand_queries_from_partial_evidence()
    ├── quality: web_research_crag_quality(top_score, unique_source_count)
    └── stop: CragRouter::should_continue(current_quality, target, hops_remaining)

3.4 CRAG Implementation (Current State)

CragRouter::expand_queries_from_partial_evidence() implements three expansion strategies:

1. Concept extraction — regex ([A-Z][a-z]{3,}...) on top-5 hit snippets → {original} {concept} queries
2. Contradiction follow-up — if hit.potential_contradiction == true → {original} conflicting evidence source comparison
3. Weak evidence — if avg_score < 0.55 → {original} primary source evidence + {original} independent corroborating sources
4. Fallback — if no concepts found → latest developments + detailed comparison

Critical gap: query expansion is heuristic regex, not LLM-generated. This means the expansion strategy is brittle for domains without capitalized proper nouns, misses technical jargon, and cannot infer what knowledge is actually missing from a research perspective.

3.5 Novelty Detection (Current State)

What exists:

• URL-level deduplication: HashSet<String> of visited URLs in run_multi_hop_web_research and dedupe_hits_by_url in the pipeline
• Source authority scoring: .gov, .edu, arxiv.org, docs.rs, github.com get score boosts
• Contradiction flag: HybridSearchHit.potential_contradiction — set by downstream heuristics
• Citation diversity: evaluate_citation_diversity — counts distinct registrable domains

What is missing (the “novelty detection” gap):

• Per-hit information gain scoring — no mechanism to compute how much new semantic content a candidate hit adds to the accumulated evidence set
• Semantic deduplication — hits with the same semantic content but different URLs are not deduplicated
• Evidence coverage map — no tracking of which aspects of the query have been covered vs. remain open
• Bidirectional write-back novelty check — new synthesis outputs are not checked for novelty before promotion to SCIENTIA

3.6 OpenRouter Free Tier Integration (Current State)

The infrastructure is wired:

• virtual_models.rs defines openrouter/free and openrouter/auto as ModelSpec entries
• openrouter/free has is_free: true, rate_limit_rpm: Some(20), rate_limit_rpd: Some(50)
• cascade_for_research_stage() in cascade.rs checks openrouter_api_key().is_some() and adds OpenRouter to the cascade
• FreeTierRouter exists in selection/free_tier.rs

Gap: The research cascade defaults to input.openrouter_model (which resolves to whatever model is configured, typically a paid model). The openrouter/free router is not automatically used as the fallback when the primary model is unavailable or when the user has no credits. There is no automatic free-tier escalation path for the research pipeline specifically.

---

4. Gap Analysis — Ranked by Leverage

#	Gap	Impact	Effort	Leverage
G1	LLM-driven CRAG query expansion	🔴 High — heuristic regex misses domain nuances	Medium	Critical
G2	Cross-encoder reranking stage	🔴 High — 8-15% precision lift in RAG benchmarks	Medium	Critical
G3	Novelty/information-gain scoring	🟠 High — redundant evidence bloats context	Medium	High
G4	Free-tier cascade fallback	🟠 High — users without API credits get no research	Low	High
G5	Confidence gate multi-signal fusion	🟠 High — routing tier stub reduces DeepResearch activation	Medium	High
G6	Synthesis token budget	🟡 Medium — 1,800 max_tokens truncates long-form reports	Low	Medium
G7	Semantic result caching	🟡 Medium — redundant search cost on similar queries	Medium	Medium
G8	Durable async research jobs	🟡 Medium — long research runs block synchronously	High	Medium
G9	Per-hop span observability	🟡 Low — debugging multi-hop loops is difficult	Low	Low

---

5. Implementation Plan — Full Deep Research Capability

Phase 1 (Quick Wins — ~2 weeks)

P1.1 — Free-Tier Research Cascade Fallback

Target: crates/vox-research-shim/src/selection/free_tier.rs + cascade.rs

Add an explicit free-tier cascade path that activates when:

• VOX_OPENROUTER_FREE_TIER=1 is set, OR
• The standard OpenRouter key resolves to a budget-constrained profile

The cascade order for free-tier research:

1. local Ollama/MENS (if available)
2. openrouter/free (openrouter/free router → DeepSeek R1:free, Llama-3.3-70B:free, Gemma-3-27B:free)
3. Fallback: SearXNG-only synthesis (no LLM, structured bullet output)

Specific model IDs to seed in the free_tier.rs catalog for research:

• deepseek/deepseek-r1:free — 164K context, excellent for reasoning-heavy research
• google/gemma-3-27b-it:free — strong instruction following
• meta-llama/llama-3.3-70b-instruct:free — balanced general research

Config additions to SearchPolicy:

/// Force research cascade onto free-tier models (env: VOX_RESEARCH_FREE_TIER).
pub research_free_tier_only: bool,
/// Preferred free-tier model IDs in priority order.
pub research_free_tier_model_ids: Vec<String>,

P1.2 — Synthesis Token Budget Uplift

Target: crates/vox-actor-runtime/src/llm/cascade.rs

Change ResearchStage::Synthesis max_tokens from 1_800 to 4_000 (configurable via ResearchConfig::synthesis_max_tokens, which already exists as a field — the issue is the apply_stage_defaults override in cascade.rs that always sets it to 1_800).

Remove the hard override in apply_stage_defaults for Synthesis — let ResearchConfig::synthesis_max_tokens win.

Phase 2 (Core Quality — ~4 weeks)

P2.1 — LLM-Driven CRAG Query Expansion

Target: crates/vox-search/src/crag.rs + research.rs

Replace CragRouter::expand_queries_from_partial_evidence() with an LLM-assisted version:

pub async fn expand_queries_llm(
    original_query: &str,
    hits: &[HybridSearchHit],
    coverage_so_far: &str,
    policy: &SearchPolicy,
) -> Vec<String>

LLM prompt (Synthesis-tier cascade, temperature=0.2):

You are a research gap analyst. Given the original research question and the evidence
collected so far, identify 2–4 specific follow-up queries that would cover the most
important missing aspects.

Original question: {original_query}
Evidence collected: {top_5_snippets}

Output only valid JSON: {"followup_queries": ["...", "..."]}
Prioritize: factual gaps, underrepresented sources, contradictions to resolve.

Fallback: if LLM cascade fails, use the existing heuristic regex expansion.

P2.2 — Novelty / Information-Gain Scoring

Target: crates/vox-search/src/ (new file: novelty.rs)

Implement a semantic novelty scorer for retrieved hits:

pub struct NoveltyScorer {
    /// Accumulated evidence fingerprints (set of normalized token n-grams).
    seen_ngrams: HashSet<u64>,
}

impl NoveltyScorer {
    /// Score 0.0–1.0: fraction of this hit's n-gram content not seen before.
    pub fn score_novelty(&mut self, content: &str) -> f64 { ... }

    /// Accept a hit (add its n-grams to the seen set).
    pub fn accept(&mut self, content: &str) { ... }
}

Algorithm: 4-gram shingling over tokens, FNV1a hashes, novelty = |new_hashes| / |total_hashes|. Hits below policy threshold (e.g., novelty_min_score: 0.15) are filtered from the synthesis context but retained in diagnostics.

Integration points:

1. run_multi_hop_web_research() — filter hop hits by novelty before accumulating
2. gather_web_hits_for_plan() — cross-subquery deduplication
3. synthesize_answer_with_llm() — sort synthesis context by novelty DESC before truncating
4. SCIENTIA write-back — check novelty of new finding candidates vs. existing corpus

SearchPolicy additions:

pub novelty_scoring_enabled: bool,
pub novelty_min_score: f64,  // default: 0.15

P2.3 — Confidence Gate Multi-Signal Fusion

Target: crates/vox-research-shim/src/research/gate.rs

Replace the Phase 0a citation-count stub with a multi-signal fusion function:

let citation_score = (citation_count / min_citations_for_full_score).clamp(0.0, 1.0);
let claim_support_score = supported_claims / total_claims;  // 0.5 if no claims
let diversity_score = (distinct_domains / min_domains).clamp(0.0, 1.0);
let retrieval_score = if no_retrieval_hits { 0.0 } else { 1.0 };

let score = citation_score * 0.35
    + claim_support_score * 0.30
    + diversity_score * 0.20
    + retrieval_score * 0.15;

This ensures the DeepResearch routing tier activates on weak multi-signal evidence, not just few citations.

Phase 3 (Advanced Quality — ~6 weeks)

P3.1 — Cross-Encoder Reranking

Target: crates/vox-search/src/reranker.rs (new file)

Add an optional reranking stage using a local cross-encoder model via vox-plugin-mens-candle-*:

pub async fn rerank_hits(
    query: &str,
    hits: &mut Vec<HybridSearchHit>,
    policy: &SearchPolicy,
) { ... }

Model candidates:

• cross-encoder/ms-marco-MiniLM-L-6-v2 (fast, 6-layer, CPU-viable)
• BAAI/bge-reranker-v2-m3 (better quality, multilingual)

Wire as a post-retrieval step in execute_search_plan() when policy.reranking_enabled = true.

SearchPolicy additions:

pub reranking_enabled: bool,
pub reranking_model_id: String,  // default: "cross-encoder/ms-marco-MiniLM-L-6-v2"
pub reranking_top_k: usize,      // default: 20
pub reranking_candidate_k: usize, // default: 200

P3.2 — Deep Research Skill

Create crates/vox-plugin-skill-deep-research/deep-research.skill.md:

name = "skill-deep-research"
description = "Full autonomous deep research: decomposes a topic into sub-queries,
retrieves from web and local corpora using hybrid search, iterates with CRAG gap-filling,
verifies claims against evidence, synthesizes a cited report, and scores it with LLM-as-judge."
[metadata]
"vox-id" = "vox.deep-research"
"vox-category" = "research"
"vox-tools" = ["vox_research_run", "vox_research_status", "vox_research_export"]

P3.3 — Semantic Result Caching

Upgrade pipeline_cache.rs from key-hash TTL cache to intent-based semantic cache:

• Store result embeddings alongside TTL entries
• On cache lookup: cosine-similarity check against stored embeddings (threshold: 0.92)
• Reduces redundant LLM costs for similar research queries by 60-70%

Phase 4 (Production Hardening — ongoing)

• P4.1 — Durable async research jobs via vox-orchestrator-queue
• P4.2 — Per-hop span tracing (research.hop.{n}.{metric})
• P4.3 — Citation diversity enforcement gate (currently computed but not enforced)

---

6. OpenRouter Free Tier — Zero-Cost Deep Research Strategy

6.1 Recommended Free Research Configuration

.env

VOX_OPENROUTER_API_KEY=<your-key>
VOX_RESEARCH_FREE_TIER=1
VOX_RESEARCH_FREE_TIER_MODELS=deepseek/deepseek-r1:free,google/gemma-3-27b-it:free,meta-llama/llama-3.3-70b-instruct:free
VOX_SEARXNG_URL=http://localhost:8080   # self-hosted SearXNG (free)
VOX_SEARCH_DDG_FALLBACK=1

6.2 Free Tier Pipeline Topology

Query
  ↓ Planner (deepseek/deepseek-r1:free, 164K ctx, reasoning)
  ↓ 3-6 subqueries
  ↓ Retrieval (SearXNG:free → DuckDuckGo:free → Tavily:optional)
  ↓ CRAG multi-hop (LLM-driven, gemma-3-27b-it:free)
  ↓ Claim extraction (llama-3.3-70b:free)
  ↓ Verification (deepseek-r1:free, temp=0.0)
  ↓ Synthesis (llama-3.3-70b:free, 4K tokens)
  ↓ Judge (gemma-3-27b-it:free, temp=0.0)
  ↓ SCIENTIA promotion (if quality ≥ 50)

Rate limit math: A single research run makes ~7–13 LLM calls. At 200 RPD, this supports 15–28 full research runs per day on free tier alone. DeepSeek R1:free has a tighter 50 RPD limit — use for high-priority stages (planner, verifier) only.

6.3 Cost-Optimized Paid Configuration

Stage	Model	Rationale
Planner	deepseek/deepseek-v4-flash	Cheapest reasoning with JSON output
Claim extraction	deepseek/deepseek-v4-flash	Speed + cost
Verification	deepseek/deepseek-r1	CoT reasoning for NLI
Synthesis	anthropic/claude-sonnet-4.6 (with caching)	Best long-form; 80%+ cost cut with prompt caching
Judge	deepseek/deepseek-v4-flash	Fast structured scoring

---

7. Novelty Detection — Detailed Design

7.1 Why Novelty Matters

Without novelty filtering, the synthesis context is padded with redundant content (e.g., 8 blog posts paraphrasing the same paper). This wastes synthesis context window, dilutes attention signal on new evidence, and inflates citation count without increasing actual coverage.

7.2 N-Gram Shingling Approach

fn shingling_hashes(content: &str, n: usize) -> Vec<u64> {
    let chars: Vec<char> = content.chars().collect();
    chars.windows(n)
        .map(|window| fnv1a_hash(&window.iter().collect::<String>()))
        .collect()
}

// novelty_score = |new_hashes| / |total_hashes|

Threshold guidance:

• novelty_min_score = 0.15 → accept hits where ≥15% of content is new (recommended default)
• 0.05 for deep technical queries where partial overlap is acceptable
• 0.30 for broad surveys to enforce source diversity

7.3 Bidirectional Write-Back (SCIENTIA Integration)

When a new research result is about to be promoted to scientia_finding_candidates, check its novelty against the existing corpus using the same shingling approach. Only promote if novelty ≥ 0.25 vs. the last 10 promoted findings in the same domain.

---

8. Existing Capability — What Works Right Now

The following deep research capabilities are production-ready today with no code changes:

1. vox research run CLI command — dispatches run_research_with_context_and_session()
2. vox_research_run MCP tool — orchestrator-MCP bridge
3. Multi-hop CRAG loop — run_multi_hop_web_research() with quality-gated stop condition
4. Hybrid BM25 + vector search — MemorySearchEngine with temporal decay and status boosting
5. RRF fusion — cross-corpus Reciprocal Rank Fusion with configurable k
6. SearXNG + DDG + Tavily web search — policy-gated, authority-boosted deduplication
7. Tavily /research deep endpoint — synthesis-quality results with answer + sources
8. Full-page HTML extraction — web-scrape feature flag, text density filtering
9. LLM query decomposition — 3–12 sub-queries with JSON schema output
10. Claim extraction + NLI verification — LLM cascade with evidence spans
11. LLM-as-judge quality scoring — 0–100 rubric (factual accuracy + citation density + coverage)
12. CoVE self-verification — consistency check before synthesis
13. Citation auditing — claim-to-source support mapping
14. SCIENTIA finding promotion — quality ≥ 50 + supported claims → scientia_finding_candidates
15. MENS training pair generation — high-confidence results → training data
16. Session tracking — research_sessions DB table with staged status updates
17. Progress callbacks — streaming progress to GUI / MCP

---

9. Prior Art Comparison — Consolidated Matrix

Feature	Gemini Deep Research	OpenClaw	Vox Current	Vox After Roadmap
Query decomposition	✅ LLM	✅ LLM	✅ LLM (cascade)	✅ Same
Multi-hop retrieval	✅ Iterative	✅ ~5 rounds	✅ Policy-gated hops	✅ Same
CRAG-style gap-fill	✅ Internal	✅ Heuristic	⚠️ Heuristic regex	✅ LLM-driven (P2.1)
Hybrid search (BM25+vec)	Internal	Varies	✅ Shipped	✅ + Reranking (P3.1)
Web search	✅ Google	Plugin	✅ SearXNG/DDG/Tavily	✅ Same
Novelty filtering	✅ Internal	Partial	❌ URL dedup only	✅ N-gram scoring (P2.2)
Cross-encoder reranking	✅ Internal	Partial	❌	✅ P3.1
Claim verification (NLI)	Internal	Partial	✅ LLM cascade	✅ Same
LLM-as-judge	✅ Internal	Partial	✅ Shipped	✅ Same
Free tier	❌ (Gemini caps)	N/A	⚠️ Infra exists, not wired	✅ P1.1
Local/private corpus	N/A	N/A	✅ Full hybrid	✅ Same
SCIENTIA promotion	N/A	N/A	✅ Unique to Vox	✅ Enhanced (P2.3)
Async durable jobs	✅ Gemini async	Varies	⚠️ Session tracked, not queued	✅ P4.1
Semantic caching	✅ Internal	N/A	⚠️ TTL only	✅ P3.3

---

10. Priority Recommendations

Immediate (this sprint)

1. P1.1 — Wire openrouter/free cascade fallback. Single config flag VOX_RESEARCH_FREE_TIER=1. Zero breaking changes.
2. P1.2 — Fix synthesis max_tokens: remove hard override in cascade.rs:apply_stage_defaults for Synthesis, let ResearchConfig::synthesis_max_tokens win.

Next sprint

3. P2.1 — LLM-driven CRAG expansion in crag.rs with heuristic fallback. Highest leverage for quality.
4. P2.2 — novelty.rs n-gram shingling. Wire into run_multi_hop_web_research and synthesis context ranking.

Following sprint

5. P2.3 — Replace gate.rs stub with multi-signal fusion.
6. P3.1 — Cross-encoder reranking using local MENS/Candle backend.

Medium term

7. P3.2 — Deep research skill YAML.
8. P3.3 — Semantic cache upgrade.
9. P4.1 — Async research job queue.
10. P4.2 — Per-hop span tracing.

---

11. Architectural Invariants to Preserve

Per AGENTS.md policy:

• All LLM calls MUST go through vox_actor_runtime::llm facade
• New free-tier model IDs must be added as ModelSpec entries in the registry, scored on the same axes (latency, cost, quality, locality)
• CRAG expansion LLM calls MUST use cascade_with_optional_manual / cascade_for_research_stage pattern
• Novelty scorer state MUST NOT be persisted across research runs (per-session working memory only)
• All new SearchPolicy fields MUST be mirrored with VOX_SEARCH_* environment variable overrides

---

12. Source References

Component	File
Full search stack	crates/vox-search/src/
Research pipeline	crates/vox-research-shim/src/research/orchestrator/
LLM cascade	crates/vox-actor-runtime/src/llm/cascade.rs
Model selection	crates/vox-research-shim/src/selection/
SCIENTIA integration	crates/vox-scientia/src/
Prior art matrix	docs/src/architecture/deep-research-prior-art-and-vox-roadmap-2026.md
SCIENTIA handoff	docs/src/architecture/scientia-automated-research-handoff-2026-06-16.md
Search retrieval SSOT	docs/src/architecture/search-retrieval-ssot-2026.md