RAG Engineer for Legal AI
Budget: ₹400 – ₹750 INR
I am building an enterprise-grade Legal AI SaaS platform that must reliably answer region-specific legal questions across all of India, with an initial focus on Corporate Law and Intellectual Property Law. The core of the product is a Retrieval-Augmented Generation (RAG) pipeline that draws exclusively from verified Indian Supreme Court and High Court judgments, returns inline citations, and resists hallucination.
What I need now is a senior, hands-on AI engineer who has already shipped production RAG systems. Your first milestone will be to design and implement the end-to-end retrieval and generation stack—ingest, indexing, querying, ranking, and grounded response generation—so that lawyers receive authoritative, citation-rich answers in real time.
Key expectations
• Build a scalable ingestion workflow for judgments (PDF, HTML, scanned orders) and enrich them with metadata suitable for vector and hybrid indexes.
• Architect the retrieval layer (elastic search, vector store, or hybrid) paired with a proven reranking strategy to maximise factual precision.
• Implement generation that embeds citations directly in the text and flags low-confidence answers.
• Establish automated evaluation to track accuracy, citation coverage, and hallucination rate.
• Expose a clean API that other micro-services (research and drafting modules will follow) can consume.
Acceptance criteria for this phase
– For a curated test set of 500 corporate and IP queries, at least 90 % of returned answers must cite the correct paragraph or page number.
– Hallucination rate (unsupported factual statements) must remain below 5 %.
– Latency under 3 s per query on commodity GPU-backed instances.
If you have demonstrable experience turning complex legal or technical corpora into trustworthy RAG products—and can speak in detail about retrieval strategies, prompt engineering, and evaluation—let’s talk.
What I need now is a senior, hands-on AI engineer who has already shipped production RAG systems. Your first milestone will be to design and implement the end-to-end retrieval and generation stack—ingest, indexing, querying, ranking, and grounded response generation—so that lawyers receive authoritative, citation-rich answers in real time.
Key expectations
• Build a scalable ingestion workflow for judgments (PDF, HTML, scanned orders) and enrich them with metadata suitable for vector and hybrid indexes.
• Architect the retrieval layer (elastic search, vector store, or hybrid) paired with a proven reranking strategy to maximise factual precision.
• Implement generation that embeds citations directly in the text and flags low-confidence answers.
• Establish automated evaluation to track accuracy, citation coverage, and hallucination rate.
• Expose a clean API that other micro-services (research and drafting modules will follow) can consume.
Acceptance criteria for this phase
– For a curated test set of 500 corporate and IP queries, at least 90 % of returned answers must cite the correct paragraph or page number.
– Hallucination rate (unsupported factual statements) must remain below 5 %.
– Latency under 3 s per query on commodity GPU-backed instances.
If you have demonstrable experience turning complex legal or technical corpora into trustworthy RAG products—and can speak in detail about retrieval strategies, prompt engineering, and evaluation—let’s talk.