Sahayak is a retrieval-augmented generation (RAG) legal assistant built by a developer named Divyansh for the PromptWars: Virtual (Exclusive Edition) hackathon. Its author says it answers tenancy, consumer-rights and contract questions, summarizes uploaded PDFs in plain language, and accepts voice questions. Its defining design choice is that when no matching context is found, it is meant to say so instead of guessing. That is a sensible intention. It is not yet proof of reliability: everything below about Sahayak is the author’s own account, and no independent legal-accuracy, calibration or refusal benchmark for it has been published that I can point to.
What Sahayak offers
The author describes three user-facing paths:
- Ask: questions on tenancy, consumer rights and contracts, answered from indexed context. If nothing relevant is retrieved, the assistant is supposed to say it cannot confirm the answer.
- Upload: native or scanned PDFs, followed by a plain-language summary covering the document type, obligations, and points worth double-checking.
- Voice: spoken questions transcribed with Whisper, then answered under the same context-grounded approach.
The author links a live demo and an MIT-licensed code repository. Those links show the stated project context; they do not verify that the demo is currently running or that the repository’s license matches the description.
](https://m.media-amazon.com/images/I/515uFKS7D4L._SL160_.jpg)
As an Amazon Associate I earn from qualifying purchases.
How it reportedly works
| Layer | Reported technology |
|---|---|
| Frontend | React and Vite, talking to the backend over REST |
| Backend | FastAPI |
| Speech-to-text | Groq Whisper |
| PDF extraction and OCR | PyMuPDF and pytesseract |
| Embeddings | sentence-transformers, all-MiniLM-L6-v2 |
| Vector search | ChromaDB |
| Answer generation | Groq LLM API, with openai/gpt-oss-120b named in the stack table |
| App data | SQLite |
The answering path, as described, retrieves relevant chunks, joins them into a context block, and sends one chat-completion request. The response returns the answer along with source names and retrieval distances. Showing sources and distances is useful because a reader can see what the answer leaned on. These are details from the author’s write-up, not a review of the deployed code.
Free tools Windows power users keep installed
One-click scans. No signup required.
Where “I don’t know” comes from
In a RAG system, refusal usually rests on two things: a retrieval step that returns nothing sufficiently relevant, and an instruction telling the model to answer only from the supplied context. Sahayak is described as using that pattern. The weak points are worth knowing:
#1 Best Overall
- Retrieval can succeed on the wrong material. A chunk can be semantically close to a question yet be out of date, from another jurisdiction, or not authoritative. Retrieval alone does not establish that a source is current or legally applicable.
- Retrieval can miss the right material. The assistant may then refuse a question it could have answered, which is a cost, not a safe default.
- Prompts are not guarantees. An instruction to stay within context reduces guessing but does not ensure every generated legal statement is supported by the retrieved text.
The author lists more jurisdiction-specific templates as future work, which suggests jurisdiction coverage is still limited. Users should treat any answer as tied to whatever material was indexed, and confirm which legal system that material covers.
What research says about teaching models to abstain
A 2024 paper by Qinyuan Cheng and colleagues studies whether AI assistants can recognize questions they do not know and say so. The authors report that models can be made more likely to refuse unknown questions. They also document a trade-off: supervised fine-tuning can cause wrongful refusals of questions the model could have answered, while preference-aware optimization can reduce some of that over-caution.
Rank #2
One headline figure from that paper: an aligned Llama-2-7b-chat could tell whether it knew the answer for up to 78.96% of questions in the authors’ test set. That is a TriviaQA-derived open-domain setup with a specific model and alignment method. It is not a Sahayak score, not a measure of legal accuracy, and not evidence that Sahayak’s refusals are dependable.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesThe lesson for a legal assistant is that abstention has a benefit and a price. Refusing unsupported answers can limit hallucination, but too much refusal makes a tool useless. A credible evaluation would measure both correct answering and appropriate refusal, on the specific legal sources, jurisdictions and document types the product claims to handle. That evaluation standard is my inference from the paper’s trade-off, not something reported for Sahayak.
Security and engineering claims
The author reports these controls, none of which has been independently audited as far as the published material shows:
- Uploaded documents and retrieved text are treated as untrusted data, not instructions, as a prompt-injection safeguard.
- Upload bytes are validated with
python-magic. - Query, upload and voice endpoints are rate-limited with
slowapi. - Production responses set CSP, X-Frame-Options and HSTS headers.
- CI workflows run linting, pytest, Bandit, Gitleaks, pip-audit and axe-core, with deployment to Render and Vercel on merges to the main branch.
The author also names persisted chat sessions and more jurisdiction-specific templates as future work, and says multi-language support is deferred. Together, that describes an early-stage project with openly acknowledged gaps.
Rank #4
Should you rely on it for a lease or contract?
Use it as a reading aid, not as advice. Reasonable uses are getting a plain-language overview of a tenancy agreement, spotting clauses to ask about, and drafting questions for a lawyer or consumer-help body. Poor uses are deciding whether to sign, withhold rent, or file a claim based on its answer alone.
Recommended Free Tools
Practical checks when using any tool of this kind:
- Look at the cited source names and ask whether they match your jurisdiction and the current law.
- Treat a confident answer with no visible supporting source as unverified.
- Treat “I don’t know” as a prompt to consult a professional, not as a sign that no answer exists.
- Avoid uploading documents with sensitive personal details unless you are comfortable with how the hosted service handles them; no independent privacy review is available.
What is and isn’t established
Established: the project exists as described by its author, with a stated architecture and a design goal of grounded, bounded answers. Not established: legal accuracy, jurisdiction coverage, how well its uncertainty is calibrated, false-refusal rate, or real-world safety. No comparison with other legal assistants is supported, so no ranking is offered here.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
