Savvy: An AI Assistant That Whispers Prompts During Meetings, a Local-First Solution Built on Your Private Documents

In an era where video meetings have become the norm in professional life, being well-prepared and composed during critical conversations remains a pain point for many professionals. Savvy, a macOS application that recently landed on the Product Hunt leaderboard (ranked #14 with 107 upvotes), offers an ingeniously crafted solution: it's an AI assistant that "whispers" what to say during meetings — and it's powered entirely by your own documents, not the open internet.

Core Positioning: A Knowledge Engine Built on Private Documents

Savvy's most distinctive feature is that its knowledge source is strictly limited to the user's own documents, rather than pulling information from the public internet. This design philosophy directly determines its use cases and trust boundaries.

The workflow is remarkably straightforward: simply point Savvy at a folder, and it builds a versioned brief for each client. This means that as your materials are updated, the briefs iterate and evolve, forming a traceable knowledge trail. For consultants, sales professionals, customer success managers, and others who frequently interact with different clients, this "organized by client, updatable anytime" knowledge structure is far more efficient than rummaging through scattered documents.

From a technical standpoint, Savvy's capability is essentially a textbook application of RAG (Retrieval-Augmented Generation) architecture. The system first segments user documents into semantic chunks, converts them into high-dimensional vectors using an embedding model, and stores them in a local vector database. When a relevant question arises during a meeting, the system similarly converts the question into a vector, retrieves the most relevant document chunks using algorithms like cosine similarity, and then passes those chunks as context to a large language model to generate responses. This "retrieve first, then generate" pipeline ensures the model's output is grounded in actual evidence rather than fabricated from thin air. More importantly, the knowledge base can be updated at any time by re-indexing documents, without the need for expensive fine-tuning.

Compared to the many meeting assistants on the market that rely on general-purpose LLMs to cast a wide net with their answers, Savvy's choice to "root itself in private materials" sacrifices breadth of knowledge in exchange for relevance and trustworthiness — which is often far more critical in formal business settings. The AI meeting assistant space is intensely competitive, with mainstream products falling roughly into three categories: meeting transcription and summarization tools like Otter.ai and Fireflies.ai, focused on post-meeting content organization; platform-level assistants like Microsoft Copilot (integrated into Teams) and Google Gemini (integrated into Meet), leveraging general-purpose LLMs for broad capabilities; and vertical tools targeting sales scenarios like Grain and Chorus.ai (acquired by ZoomInfo), specializing in customer conversation analysis. Savvy's uniqueness lies in the fact that it neither does post-meeting summaries nor relies on the generalized knowledge of universal LLMs. Instead, it focuses on real-time knowledge retrieval and prompting during meetings, with knowledge sources entirely under user control. This "narrow but deep" positioning avoids head-on competition with major tech companies, though it also means Savvy needs to build strong enough word-of-mouth and user stickiness within its specific target audience.

Restrained Three-Trigger Interaction: Speaking Up Only at Critical Moments

One of the most fascinating aspects of Savvy's product philosophy is its restraint. During a meeting, it doesn't chatter incessantly. Instead, it stays quiet and only "speaks up" under three specific conditions:

• The other party asks a question your brief can answer — relevant information surfaces proactively;
• Someone crosses a "red line" you've pre-set — triggering alerts on sensitive topics or non-negotiable terms you've flagged;
• You manually press the "Advice" button — summoning prompts on demand.

This "three-trigger" mechanism addresses a common pain point with AI assistants: information overload and unnecessary interruptions. A tool that constantly pops up prompts during a meeting actually fragments attention and disrupts the natural rhythm of conversation. By handing more control back to the user and limiting automatic prompts to high-value moments, Savvy attempts to strike a balance between "being useful" and "not being intrusive."

Every Card Cites Its Source

It's worth emphasizing that every prompt Savvy delivers (presented as a "card") cites its source. This means users can immediately see which document and which version a suggestion came from, enabling quick assessment of its credibility.

At a time when AI hallucination remains a pervasive problem, this "traceable" design not only enhances practical utility but is also a responsible approach for professional scenarios. AI hallucination refers to the phenomenon where large language models generate responses that sound confident and fluent but are factually incorrect, fabricated, or inconsistent with the input data. The root cause lies in the fact that LLMs are fundamentally probabilistic text generation systems — they predict the next most likely token rather than retrieving facts from a structured database. In business meeting scenarios, the risk of hallucination is particularly dangerous: if an AI assistant fabricates a contract clause, invents customer history data, or provides a non-existent quote, the consequences range from damaged professional credibility to potential legal disputes. By strictly limiting knowledge sources to user-owned private documents and citing the origin of every suggestion, Savvy uses deterministic document retrieval to constrain probabilistic model output, significantly reducing the likelihood of hallucination.

Privacy-First Local Architecture

For professionals handling sensitive client materials, privacy and data security are virtually non-negotiable. Savvy makes clear architectural commitments in this regard:

Documents, indexes, and transcripts all remain on your Mac locally — only "extracted snippets" and "audio streams" leave the device. In other words, your complete knowledge base is never uploaded; only the minimum necessary information is transmitted during AI processing. This "local-first, cloud-supplemented" hybrid model represents a pragmatic trade-off between capability and privacy.

The requirement for Apple Silicon and macOS 13 or later isn't just a simple compatibility limitation — it's a critical pillar of the product's architecture. Since Apple introduced the M1 chip in 2020, its custom silicon has included a dedicated Neural Engine, with the latest M4 series delivering up to 38 TOPS (trillion operations per second) of AI inference capability. Combined with Apple's Core ML framework and Metal GPU acceleration, developers can efficiently run embedding models to generate vectors, perform real-time speech transcription (for example, based on OpenAI's open-source Whisper model), and execute semantic document retrieval — all locally. It's precisely this powerful on-device compute that enables Savvy to keep computationally intensive tasks like document indexing and audio transcription on the local machine, only transmitting the minimum necessary information to the cloud when LLM inference is needed. This achieves a technically feasible balance between AI capabilities and privacy protection. It also explains why Savvy can keep so much data local — modern Macs have enough on-device compute to handle a substantial portion of inference tasks.

Open-Source Transparency: Trust Guaranteed by the MIT License

Another major highlight of Savvy is that it's open-sourced under the MIT License. The MIT License is one of the most permissive open-source licenses available, created by the Massachusetts Institute of Technology. It allows anyone to freely use, copy, modify, merge, publish, distribute, sublicense, and sell copies of the software, with the sole requirement of retaining the original copyright notice and license text. Compared to copyleft licenses like the GPL, which require derivative works to also be open source, the MIT License imposes virtually no restrictions on commercial use.

For a tool that handles sensitive business data, open source means greater transparency: security researchers can audit the code to verify whether privacy claims hold true — for instance, confirming that data isn't being secretly transmitted to third-party servers or checking for security vulnerabilities. Enterprise users can also deploy it on internal infrastructure and customize it to their needs.

In the privacy tools space, "trust" is often difficult to establish through vendor promises alone, and auditable open-source code provides a more solid path to trust. This kind of "verifiable trust" is more convincing than the privacy pledges of closed-source software, especially when dealing with sensitive business data like client contracts and quotations. This also aligns seamlessly with Savvy's overall product values of "local-first, source-traceable."

Summary: A Vertical and Restrained Professional Meeting Assistant

From a product design perspective, Savvy represents an emerging class of AI application philosophy: rather than pursuing an all-knowing general-purpose assistant, go deep and thorough in a specific scenario. Its key decisions — binding to private documents, restrained three-trigger interaction, source citations, local processing, and open source — complement each other, collectively serving the clearly defined scenario of "professional business meetings."

Of course, as an early-stage product, it has clear limitations: the macOS and Apple Silicon requirement confines its audience to the Apple ecosystem; the "private document-based" model means its value is highly dependent on the quality and organization of the user's pre-existing materials — if documents are poorly organized, outdated, or inconsistently formatted, RAG retrieval quality will suffer accordingly. But it's precisely this decisively defined positioning that gives it a distinct identity in an increasingly homogeneous meeting assistant market.

For professionals who frequently need to draw on extensive background materials during meetings and place a high premium on data privacy, Savvy is well worth trying.