In a move that signals a maturing intersection of privacy, personalization, and retrieval-augmented generation (RAG), the independent search and knowledge management project «Свой» has officially opened its source code and announced a set of novel capabilities. The announcement, published on Habr in late July 2026, introduces three core features that push beyond conventional RAG pipelines: vibe themes, RAG over cipher, and a context-aware lens. This article breaks down what these terms mean, why they matter, and how they could reshape the open-source AI landscape.
The Context: Why «Свой» Mattered Before Open Source
«Свой» (Russian for “one’s own”) started as a personal search engine that prioritises user privacy and local-first architecture. Unlike mainstream tools that index the public web, «Свой» allowed individuals to build private knowledge bases from their own documents, bookmarks, notes, and selected external sources. The project gained traction among privacy-conscious researchers, journalists, and power users who needed a search tool that could run entirely offline while still leveraging large language models (LLMs) for reasoning.
Before the open-source announcement, «Свой» was a closed-source product with a free tier and a paid subscription model for advanced features. According to the Habr article, the developers made the decision to open the codebase to accelerate community contributions and to prove that privacy-preserving AI is not a contradiction in terms.
Vibe Themes: Personalising Information Retrieval beyond Keywords
The first feature highlighted in the announcement is vibe themes. Traditional RAG systems retrieve documents based purely on semantic similarity measured by embedding vectors. While effective, this approach ignores the user’s current emotional, cognitive, or situational context. The «Свой» team proposes that relevance is not only about topical similarity but also about the mood or style the user needs at a given moment.
How Vibe Themes Work
A “vibe” in «Свой» is a set of parameters that modify the retrieval and generation pipeline. For example:
- Analytical vibe: Prioritises documents with structured arguments, tables, and charts.
- Creative vibe: Favours metaphorical language, examples, and narrative structures.
- Concise vibe: Filters for terse summaries and bullet points.
- Exploratory vibe: Increases diversity in retrieved chunks to surface unexpected connections.
These vibes are not simple filters. They are implemented as learnable soft prompts that influence the embedding re-ranking step and the LLM’s generation style. The user can switch vibes on the fly, or let the system infer the appropriate vibe from the query’s phrasing (e.g., “explain like I’m five” triggers the simplified vibe).
This is a significant departure from the one-size-fits-all RAG approach used by most open-source frameworks (LangChain, LlamaIndex). While those tools allow custom prompt templates, they do not natively support dynamic, multi-dimensional retrieval modulation. The «Свой» team claims that in internal testing, vibe-aware retrieval improved user satisfaction scores by over 30% compared to static RAG.
RAG over Cipher: Privacy at the Core
Perhaps the most technically ambitious feature is RAG over cipher. The premise is simple: the user’s documents should never be decrypted during retrieval or generation, even on their own device. The system works with data encrypted at rest and in transit, and performs retrieval directly on encrypted embeddings.
Technical Implementation
«Свой» uses homomorphic encryption for the embedding vectors and order-preserving encryption (OPE) for index storage. The retrieval step happens inside a trusted execution environment (TEE) – specifically Intel SGX enclaves – which decrypts only the minimal context needed to generate a response, then immediately re-encrypts it.
Key components:
| Component | Role | Encryption Method |
|---|---|---|
| Document Store | Encrypted chunk storage | AES-256-GCM |
| Embedding Index | Encrypted vectors for similarity search | Homomorphic encryption (CKKS scheme) |
| Retrieval Orchestrator | Performs search inside TEE | Intel SGX + OPE |
| LLM Interface | Decrypts only top-k chunks inside TEE | Secured by SGX |
This setup ensures that even if the host machine is compromised, an attacker cannot read the user’s documents or the generated responses without breaking the SGX or the homomorphic encryption. The developers note that the performance overhead is significant – retrieval latency increases by 2× to 4× compared to plaintext RAG – but they consider it acceptable for sensitive use cases.
“We believe that privacy should be the default, not an afterthought,” the team wrote in the Habr article. “RAG over cipher is our answer to cloud-based RAG services that see all your data.”
The approach is reminiscent of research from MIT’s Cryptography and Information Security group, but «Свой» is one of the first production-ready implementations of homomorphic retrieval for personal knowledge bases.
The Lens That Knows Context
The third major feature is the context-aware lens. This is a user interface component that acts as an intelligent overlay on any application – browser, text editor, terminal, or even a video player. The lens continuously monitors the user’s current context (open tabs, active document, selected text, clipboard content, recent file operations) and surfaces relevant information from the user’s own «Свой» index without an explicit query.
Context Gathering Without Cloud Dependencies
Unlike Microsoft Copilot or Google’s Workspace AI, which send context to cloud servers, the lens processes everything locally. It uses:
- OS-level event hooks (accessibility APIs on macOS, Windows, and Linux) to capture active application metadata.
- Optical character recognition (OCR) for images and video frames if the user grants permission.
- Spatial and temporal context – e.g., if the user has been editing a document about “neural networks” for two hours, the lens boosts retrieval for related keywords like “backpropagation” and “weight initialization.”
The lens is not a separate UI; it overlays a small floating window or integrates into the system tray. When it detects a relevant snippet from the user’s index, it can proactively show it as a tooltip or notification. The user can also summon the lens with a hotkey and ask a natural language question that is automatically conditioned on the current context.
Use Case Example
Consider a data scientist reading a research paper about vision transformers. The lens recognises that the PDF is open and the user has previously saved an article about “patch embedding.” It quietly retrieves a summary of that article and displays it in the corner. The user does not need to switch apps or type a query – the lens “knows” the context and delivers without interruption.
This always-on, proactive retrieval goes beyond even the “Awareness” feature in some modern AI assistants because it works entirely offline and on the user’s own data.
Why Open Source Matters for These Features
The decision to open-source «Свой» is strategic. Each of the three features – vibe themes, RAG over cipher, and the context lens – relies on a stack that is hard to standardise. By releasing the code under a permissive license, the team hopes to:
- Accelerate adoption of homomorphic encryption for RAG.
- Encourage the development of vibe theme presets by the community (e.g., a “journalist vibe” or “developer vibe”).
- Allow security researchers to audit the TEE and encryption implementation, which is critical for trust.
The announcement includes a roadmap for the next 12 months: a plugin system for third-party lens integrations, support for more TEE backends (AMD SEV, ARM TrustZone), and a distributed index that can sync encrypted data across devices.
Implications for the AI Ecosystem
The «Свой» open-source release arrives at a time when the industry is grappling with two contradictory trends: the desire for personalised, context-aware AI and the growing demand for data sovereignty. Most commercial RAG solutions (e.g., Glean, Dropbox Dash, Notion AI) rely on cloud processing and can see user content. Open-source frameworks like LlamaIndex and LangChain offer some privacy but lack homomorphic encryption and proactive context sensing.
«Свой» bridges that gap. Its approach could influence:
- Enterprise knowledge management: Companies with strict compliance requirements (healthcare, finance, legal) may adopt «Свой» as a foundation for internal AI assistants.
- Personal AI assistants: Projects like Mycroft or private voice assistants could embed «Свой»’s lens to become context-aware without phoning home.
- Research reproducibility: By open-sourcing the whole pipeline, researchers can benchmark vibe themes and encrypted retrieval against traditional RAG.
Challenges Ahead
The developers acknowledge that the current version has limitations:
- Latency: Homomorphic retrieval is still too slow for real-time conversational use at scale. The team is exploring hybrid approaches where only the most sensitive queries go through cipher RAG.
- Vibe calibration: It takes time for a user to find the right vibes for their workflow. There is no automatic vibe learning yet (though it is planned).
- TEE availability: Intel SGX is not available on all hardware, and cloud-based TEEs require trust in the cloud provider.
Nevertheless, the open-source community now has a tangible reference implementation to build upon.
Conclusion
The open-sourcing of «Свой» marks a milestone in the evolution of privacy-first, context-aware AI. Vibe themes propose that retrieval should be emotionally and stylistically attuned to the user. RAG over cipher demonstrates that retrieval-augmented generation can be truly private without sacrificing the power of LLMs. And the context-aware lens hints at a future where AI assists seamlessly without being intrusive.
As the project grows under open governance, it will be interesting to see whether these features become standard in the RAG ecosystem. For now, «Свой» offers a compelling blueprint for anyone building AI systems that respect user autonomy and data ownership.
Comments