핵심 요구사항
- 암호화(전송·저장 중) 적용된 벡터DB 인프라 코드(Docker/Terraform)
- PDF·DOCX·텍스트 수집 및 파일 변경 시 임베딩 동기화 스크립트
- 인증·모델 토글·스트리밍 응답을 지원하는 쿼리 API/웹 UI
- 10k 문서 기준 3초 이내 쿼리 지연
- 모든 응답에 소스 파일·페이지/문단 단위 인용 포함
- 재배포·확장 가능한 명확한 문서화
Required stack
필요 기술
PythonLangChainVector DatabaseDockerTerraformOpenAI APIAnthropic Claude APIFastAPI
Project brief
프로젝트 내용
I need a developer who can wire a private vector store to either use OpenAI’s API and Anthropic’s Claude API so I can question my own directory of case law and statutes without anything leaking beyond my walls. Speed matters: queries should return in seconds, even when the document set grows. Security is non-negotiable, and the architecture must scale as I add more files or swap to bigger cloud instances.
Here’s the end-to-end flow I’m after: I drop new PDF opinions or briefs into a directory, an ingestion routine chunks and embeds them, the embeddings land in a secure vector database, and a slim UI (or endpoints I can call from my legal system) lets me ask a question. The system then composes a response—whether it is an email to opposing counsel, a client letter, or material for a formal legal brief—citing the sources it pulled from my corpus.
Essential pieces I expect from you:
• Infrastructure code (Docker, Terraform or similar) that spins up the vector database with proper encryption at rest and in transit
• Ingestion script that handles common legal file types (PDF, DOCX, text) and keeps embeddings in sync when I update or delete a file
• Query API or lightweight web front end with auth, model toggle, and streaming answers
• Clear instructions so I can redeploy or extend the stack myself
Acceptance criteria
1. Query latency under 3 s on a 10 k-document test set
2. End-to-end encryption verified by independent tools (provide commands or screenshots)
3. Citations in every response linking back to the exact source file and page/paragraph
4. Successful generation of an email, a client letter, and a brief excerpt from the same prompt during demo
If you have prior work with secure RAG pipelines, let’s talk.
Here’s the end-to-end flow I’m after: I drop new PDF opinions or briefs into a directory, an ingestion routine chunks and embeds them, the embeddings land in a secure vector database, and a slim UI (or endpoints I can call from my legal system) lets me ask a question. The system then composes a response—whether it is an email to opposing counsel, a client letter, or material for a formal legal brief—citing the sources it pulled from my corpus.
Essential pieces I expect from you:
• Infrastructure code (Docker, Terraform or similar) that spins up the vector database with proper encryption at rest and in transit
• Ingestion script that handles common legal file types (PDF, DOCX, text) and keeps embeddings in sync when I update or delete a file
• Query API or lightweight web front end with auth, model toggle, and streaming answers
• Clear instructions so I can redeploy or extend the stack myself
Acceptance criteria
1. Query latency under 3 s on a 10 k-document test set
2. End-to-end encryption verified by independent tools (provide commands or screenshots)
3. Citations in every response linking back to the exact source file and page/paragraph
4. Successful generation of an email, a client letter, and a brief excerpt from the same prompt during demo
If you have prior work with secure RAG pipelines, let’s talk.
지원 기회 분석
지원 전에 볼 것
예상 산출물
- 암호화 벡터DB 인프라 코드(Docker/Terraform)
- 법률 파일 수집·임베딩 동기화 스크립트
- 인증·스트리밍 지원 쿼리 API 또는 웹 프론트엔드
- 재배포/확장 가이드 문서
매력 포인트
- 명확하고 구체적인 스펙과 수용 기준
- 성능·보안 지표가 정량적으로 정의됨
확인할 점
- 요구 범위(보안 RAG 풀스택 + IaC + 인증 + 스트리밍)에 비해 예산 $20~250이 심각하게 낮음
- 기간 명시 없음
- 통화 표기 혼동($ 기호+GBP)
Client signal
클라이언트 정보
플랫폼에서 상세 정보를 확인하세요.