Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsiTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Sahayak is a legal-assistant project whose author describes it as a retrieval-augmented AI system for questions about tenancy, consumer rights, and contracts. It is also intended to summarize uploaded PDFs and accept spoken questions. Its defining promise is bounded answering: when the retrieved material does not support an answer, it should acknowledge that rather than guess. That is a useful design goal, but the available project description does not establish that Sahayak is legally accurate or that its uncertainty judgments are reliable.
What Sahayak is described as doing
In the project author’s description, Divyansh presents Sahayak as a tool for making legal information easier to understand, not as a demonstrated substitute for a lawyer. The author describes three ways to use it:
- Ask: Submit a question about tenancy, consumer rights, or contracts. The system is intended to answer from indexed text and disclose when it cannot find supporting context.
- Upload: Provide a native or scanned PDF for a plain-language summary covering the document type, obligations, and points to double-check.
- Voice: Ask by speech; the author says Whisper transcribes the question before the answer is generated from retrieved context.
These are capabilities reported by the author, not independently verified product results. The author says the project was built for the PromptWars: Virtual (Exclusive Edition) hackathon and links to a demo and source repository. Those links describe the project’s stated context; they do not establish that the demo remains available or independently confirm the repository’s current license terms.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →How the “I don’t know” behavior is supposed to work
Sahayak is described as using retrieval-augmented generation (RAG). In the reported answering flow, it retrieves text chunks related to a question, combines them into context, and sends that context with the question in a chat-completion request. The response is described as including the answer, source names, and retrieval distances.
#1 Best Overall
That flow can make it easier to see what material influenced an answer, but retrieval is not a legal fact-check. A retrieved passage may be irrelevant, out of date, incomplete, or from a source that is not authoritative for the reader’s jurisdiction. Even a relevant passage does not prove that every legal proposition in the generated answer follows from it. Saying “I don’t know” when support is missing is therefore a useful safeguard against guessing, not evidence that the answers it does give are correct.
Reported technology and safeguards
The project article reports a React and Vite frontend communicating with a FastAPI backend over REST. It names Groq Whisper for speech transcription; PyMuPDF and pytesseract for PDF extraction and OCR; sentence-transformers with all-MiniLM-L6-v2 for embeddings; ChromaDB for vector retrieval; the Groq API for answer generation; and SQLite for app data. The article names openai/gpt-oss-120b as the Groq-served model. These are the author’s reported implementation details, not an independent inspection of the deployed system.
Rank #2
The author also reports several security measures: treating uploaded documents and retrieved text as untrusted data rather than instructions, validating upload bytes with python-magic, rate-limiting query, upload, and voice endpoints with slowapi, and setting CSP, X-Frame-Options, and HSTS headers in production. The project description says its CI workflows include linting, pytest, Bandit, Gitleaks, pip-audit, and axe-core, with deployment to Render and Vercel on merges to the main branch. No security audit or independent test results are provided, so these should be understood as reported controls rather than verified security guarantees.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →What “I don’t know” can—and cannot—tell you
Abstention means declining to answer when a system judges that it lacks enough knowledge or evidence. It can reduce unsupported answers, but a model can also refuse a question it could have answered. The tradeoff appears in general research on uncertainty expression: a 2024 study by Cheng and coauthors found that training can make models more likely to refuse questions they do not know, while some supervised fine-tuning methods also produced incorrect refusals on answerable questions. The study concerns open-domain question answering, not Sahayak or legal advice.
In one experiment, the researchers reported that aligned Llama-2-7b-chat could determine whether it knew the answers for up to 78.96% of questions in their specific TriviaQA-derived test set. That figure is not Sahayak’s score, a legal accuracy rate, or evidence that Sahayak’s refusals are dependable. It illustrates why an abstention rate alone would not be enough: an evaluation must also check whether the system answers questions it can support.
What is not established about Sahayak
The project description and the separate abstention study do not provide a Sahayak-specific benchmark for legal accuracy, correct retrieval, uncertainty calibration, appropriate refusals, or real-world safety. They also do not establish which jurisdictions or legal sources the system covers, how current those sources are, or how it performs on scanned documents in practice.
The author identifies persisted chat sessions and more jurisdiction-specific templates as future work, and says multilingual support is deferred. That makes it especially important not to assume that a general answer applies to a particular place, language, or legal situation. A system can retrieve text without proving that the text is current law or applies to the person asking.
Recommended Free Tools
How to assess a legal AI assistant like Sahayak
Before relying on a response, check whether the system identifies sources you can inspect and whether those sources are authoritative and current for the relevant jurisdiction. For a meaningful independent evaluation, testing would need to cover both answerable and unanswerable legal questions, alongside the document types and jurisdictions the tool claims to support. Useful questions include:
- Can you identify the source behind each important legal claim, and verify its authority and date?
- Does the system distinguish jurisdictions instead of presenting a local rule as universal?
- Does it abstain when evidence is missing without refusing questions its sources do support?
- Has its legal performance been independently evaluated on representative questions and documents?
- Are privacy and security practices independently documented, beyond claims about implemented controls?
- Does it clearly explain when a summary or answer needs review by a qualified professional?
For a contract or tenancy document, treat an AI summary as a reading aid: compare important obligations, dates, fees, and termination terms with the original text. Do not rely on an unsupported answer to make a legal decision.
Quick Recap
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.
