💡 Conversational voice assistant grounded in hive history

[github-actions[bot]]

Summary

Extend Broodly's voice capture from one-way logging into a true bidirectional, hands-free assistant grounded in the user's own hive history. A beekeeper standing in front of Hive 4 with sticky gloves should be able to ask "When did I last treat this hive? What did I see in the brood pattern last visit? Is this colony due for a split?" and get a precise, sourced answer in the same session. This is the natural amplification of Broodly's zero-tap differentiator.

Market Signal

• APiLOG launched in 2026 marketing an AI assistant that "learns from your hive records to provide personalized insights, answer questions, and suggest tasks" — and they explicitly tie it to voice and video capture. They are the first competitor to ship this.
• HiveHelp.AI markets AI Q&A but is not grounded in user records — it's a generic chatbot, which beekeepers correctly distrust for safety-critical decisions.
• The category is forming right now: there is a 6–12 month window before "AI assistant grounded in your own hive history" becomes table stakes the way "voice logging" did.

User Signal

• The PRD's zero-tap beeyard mantra in CLAUDE.md is the strongest single differentiator the project has — but today voice is one-way (capture only). Users in the field still need to look at the screen to see answers. That breaks the zero-tap promise.
• The Newbie persona's "decision uncertainty at high-impact moments" pain (prd.md Executive Summary) is exactly what a grounded conversational assistant addresses: "what does this brood pattern usually mean?" answered with the user's own historical context, not a generic textbook.

Technical Opportunity

• The recommendation engine already produces explainable outputs (action + rationale + confidence). Wrapping it as a tool callable by an LLM with retrieval over the user's InspectionLog, Treatment, and Recommendation records is a natural extension — not a rewrite.
• pgvector is already in the stack (architecture.md) and Vertex AI Embeddings 2.0 + Gemini are in the tech stack quick reference. The infra cost to add a query_hive_history(hive_id, question) tool is low.
• TTS for the response is the only new client capability needed; both iOS (AVSpeechSynthesizer) and Android (TextToSpeech) ship with system TTS that meets the offline-first constraint.
• Critically: the assistant grounds every answer in concrete log entries with timestamps, eliminating hallucination risk on safety-critical recommendations.

Assessment

Dimension	Score	Rationale
Feasibility	medium	Stack is in place (pgvector, Gemini, voice). Hard parts are latency budget, prompt safety, and TTS UX in noisy outdoor environments.
Impact	high	Amplifies the single most differentiated feature in the product. Every voice interaction becomes a moat.
Urgency	high	APiLOG is here; the category is forming now; first-mover defensibility decays fast.

Adversarial Review

Strongest objection: "LLMs hallucinate. Recommending the wrong treatment to a Newbie because the model invented a fact is an existential failure mode for a field-first decision-support product."

Rebuttal: The mitigation is architectural, not a prompt-engineering hope: every assistant answer must cite a specific log entry by ID and date, and any non-cited claim is rejected by a post-processing guard. If the grounding retrieval returns no relevant log, the assistant is required to say "I don't have that in your records" — never invent. This is the same safety contract the existing recommendation engine already enforces (action + rationale + confidence + fallback); the assistant just inherits it. Where the existing recommendation engine is the source of truth, the assistant calls it as a tool rather than re-reasoning.

Suggested Next Step

Run a one-week spike: build a CLI prototype that takes a hive ID and a natural-language question, retrieves relevant logs from a seeded test database via pgvector, and returns a Gemini-generated answer with mandatory citations. Measure: latency (target <3s on Cloud Run), citation accuracy (target 100% — zero ungrounded claims), and answer usefulness on a 20-question test set drawn from the brainstorming session notes.

---

Proposed by the BMAD Analyst (Mary) on 2026-04-07.
Workflow run

[github-actions[bot]]

Weekly Update

What Changed

The voice-first beekeeping space has become significantly more competitive since this idea was proposed:

• HIVESOUND ships a multi-colony continuous voice session where beekeepers mention colony names and observations are automatically assigned to the correct colony. Their voice recognition "understands the context of beekeeping in depth" and converts natural speech into structured datasets. This directly overlaps with Broodly's planned conversational voice capability.
• BeeKeeperVoice ships voice-controlled checklists with speech recognition trained on beekeeping terminology ("varroa mite count," "supersedure cell") and AI-powered health tips. iOS only currently, Android coming.
• APiLOG offers voice/video transcription with automatic extraction of key data points (brood patterns, treatments, resource levels), plus an AI assistant that "answers questions and generates personalized tasks based on your hive's real data and history." This is close to the RAG-based conversational Q&A described in this proposal.
• HiveSense ships on-device Whisper AI for offline voice transcription — a critical capability gap for Broodly's cloud-dependent architecture.
• Broodly's PRD already includes "Conversational inspection Q&A" (FR post-MVP, Phase 2) using RAG over curated domain knowledge combined with user hive context.

Updated Assessment

Dimension	Previous	Current	Delta
Feasibility	—	med	→ (first assessment; proven by competitors, but RAG-based approach adds complexity over simple voice logging)
Impact	—	high	→ (validated by market — 4+ competitors now ship voice features, confirming demand)
Urgency	—	high	↑ (from uncontested to crowded — HIVESOUND and APiLOG ship features that overlap with this proposal's scope)

Recommendation

Advance with differentiation focus — Voice logging alone is no longer a differentiator. Broodly's unique angle is the conversational assistant grounded in hive history — not just recording observations but engaging in a two-way dialogue that uses the recommendation engine and historical context to guide the conversation. Key differentiation points: (1) confidence-scored responses that distinguish certain from uncertain guidance, (2) hive-history-grounded answers rather than generic knowledge, (3) proactive prompting based on what the recommendation engine thinks should be checked. The risk is competing on voice UX alone against established voice-first apps — the value must come from the intelligence behind the voice, not the voice interface itself.

[github-actions[bot]]

Weekly Update

What Changed

Competitive and technology shifts relevant to conversational voice assistants:

• HiveTracks AI Voice: HiveTracks v3.0 shipped an AI voice feature that lets users "talk through inspections and have data saved automatically." Cloud-dependent (not offline). They are actively developing context-aware chat and tailored recommendations based on hive history — this is the closest competitive approach to Broodly's conversational voice assistant concept.
• Gemini 3.1 Pro: 1M-token multimodal context window. This means an entire season of hive history (inspection observations, recommendations, photos, voice transcripts) can be loaded as context for a single conversational turn. Enables deeply grounded, history-aware voice interactions without complex RAG infrastructure.
• Connection AI Badge (agriculture): Kickstarter June 2026 — an AI voice badge for hands-free farm coordination. Demonstrates growing market validation for voice-first agricultural tools.
• Herd Advisor (livestock): Launched at CattleCon 2026 — voice-to-record for cattle management. Validates the hands-free field logging pattern across agriculture verticals.
• PRD FR12b2 (Live Mode): The PRD added Live Mode — accessible from a persistent EQ icon in the app bar, it verbally summarizes all current context via TTS (weather, bloom, scale trends, pending actions, alerts). This is a first step toward the conversational assistant vision.

Updated Assessment

Dimension	Previous	Current	Delta
Feasibility	med	high	↑ (Gemini 3.1 Pro 1M-token context eliminates need for complex RAG — load full hive history directly)
Impact	high	high	→ (HiveTracks entering the space validates demand; Broodly's offline + grounded approach differentiates)
Urgency	med	high	↑ (HiveTracks is shipping AI voice features now; execution timing matters)

Recommendation

Advance. HiveTracks' cloud-dependent AI voice validates market demand but creates an opening for Broodly's offline-capable, confidence-scored alternative. Gemini 3.1 Pro's 1M-token context dramatically simplifies the technical approach: instead of building complex RAG over curated domain knowledge, load the user's complete hive history plus regional context directly into Gemini's context window. Combined with Live Mode (FR12b2), this creates a natural escalation path: Live Mode → conversational Q&A → full conversational voice assistant.

Sources:

• HiveTracks — From Beekeeping to Environmental Monitoring
• Connection — AI Voice Badge for Farms
• Gemini 3.1 Pro Model Card
• Herd Advisor — Voice-to-Record for Cattle