DocMind

Document intelligence API powered by Gemini and Vertex AI RAG Engine.

Upload documents, index them into a Vertex RAG corpus, retrieve relevant chunks, and answer questions with grounded citations.

Features

• Vertex AI RAG corpus management (auto-create or reuse existing corpus)
• File ingestion to Vertex RAG (upload_file) with chunking controls
• Semantic retrieval via rag.retrieval_query
• Vertex-native LLM reranking (RagRetrievalConfig.ranking.llm_ranker)
• Grounded answer generation with Gemini Flash and optional Pro fallback
• Auto-ingestion from docs/ with file watcher
• Optional API key auth and local/GCS storage

Prerequisites

• Python 3.12–3.13
• GCP project with Vertex AI API enabled
• Application Default Credentials:

gcloud auth application-default login

Quick Start

git clone https://github.com/EhsanulHaqueSiam/docmind.git && cd docmind
uv sync
cp .env.example .env

Set at least:

GCP_PROJECT_ID=your-project-id
GCP_LOCATION=us-central1

Run:

uv run uvicorn src.main:app --reload --host 0.0.0.0 --port 8000

Open:

• http://localhost:8000 (documentation page)
• http://localhost:8000/docs (Swagger UI)

API Endpoints

Method	Endpoint	Description
POST	/api/v1/query	Ask a question with RAG grounding
POST	/api/v1/search	Retrieve relevant chunks (with optional rerank)
POST	/api/v1/documents/upload	Upload + ingest a document
GET	/api/v1/documents	List ingested documents
DELETE	/api/v1/documents/{doc_id}	Delete an ingested document
POST	/api/v1/documents/ingest	Re-ingest all files from docs/
GET	/api/v1/health	Health check (Vertex RAG connectivity)
GET	/api/v1/stats	Vertex RAG corpus/file stats

Configuration

Variable	Default	Description
GCP_PROJECT_ID	—	Google Cloud project ID
GCP_LOCATION	us-central1	Vertex AI region
GEMINI_FLASH_MODEL	gemini-2.5-flash	Default answer model
GEMINI_PRO_MODEL	gemini-2.5-pro	Fallback answer model
VERTEX_RAG_CORPUS_NAME	—	Existing corpus resource name/ID to reuse
VERTEX_RAG_CORPUS_DISPLAY_NAME	docmind-corpus	Corpus display name when auto-creating
VERTEX_RAG_CORPUS_DESCRIPTION	DocMind managed corpus	Corpus description
VERTEX_RAG_EMBEDDING_MODEL	publishers/google/models/text-embedding-005	Embedding model for corpus backend
VERTEX_RAG_CHUNK_SIZE	512	Chunk size for upload transformations
VERTEX_RAG_CHUNK_OVERLAP	100	Chunk overlap
VERTEX_RAG_VECTOR_DISTANCE_THRESHOLD	—	Optional retrieval filter
VERTEX_RAG_ENABLE_LLM_RERANK	true	Enable Vertex LLM reranking
VERTEX_RAG_RERANKER_MODEL	gemini-2.5-flash	LLM reranker model
TOP_K	10	Retrieval candidates
RERANK_TOP_K	5	Final chunks used
STORAGE_MODE	local	local or gcs
DOCS_DIRECTORY	docs	Watched docs directory
GCS_BUCKET	—	Required when STORAGE_MODE=gcs
API_KEY	—	If set, require X-API-Key
MAX_UPLOAD_SIZE_MB	100	Upload size limit
WATCHER_DEBOUNCE_SECONDS	2.0	File watcher debounce

Architecture

Upload file
  -> save to storage
  -> vertexai.rag.upload_file(corpus, chunking config)

Search
  -> vertexai.rag.retrieval_query(top_k, optional llm ranker)

Query
  -> search_with_rerank()
  -> build context with retrieved chunks
  -> Gemini generate content (Flash -> optional Pro fallback)

Deploy to Cloud Run

• Build and push image.
• Set env/secrets (project ID, region, bucket, optional API key).
• Deploy deploy/cloudrun.yaml.

Notes

Vertex RAG supported regions and availability can change. Check official Vertex AI RAG docs before production rollout.