LLM Document Analysis Pipeline

Microservice that classifies, summarizes and extracts structured JSON data from unstructured PDF and text documents using Claude (Anthropic API) and LangChain, plus a RAG pipeline backed by a FAISS vector store with Voyage AI embeddings. Exposed via async FastAPI endpoints and containerized with Docker Compose.

Stack

Python · LangChain · Anthropic API · Voyage AI · FastAPI · FAISS · Docker

API

Endpoint	Description
POST /documents/analyze	Classify + summarize + extract in one call (runs concurrently)
POST /documents/classify	Classify document into a category (invoice, contract, report, …)
POST /documents/summarize	Summary in the document's language
POST /documents/extract	Structured metadata as JSON (title, date, author, entities, keywords)
POST /rag/ingest	Chunk a document and add it to the FAISS index
POST /rag/query	Ask a question over the ingested documents (RAG)
GET /health	Health check

Document endpoints accept multipart/form-data uploads (PDF or plain text). Interactive docs at http://localhost:8000/docs.

Quick start

cp .env.example .env   # add your Anthropic and Voyage AI API keys
docker compose up --build

Example:

curl -F "file=@report.pdf" http://localhost:8000/documents/analyze
curl -F "file=@report.pdf" http://localhost:8000/rag/ingest
curl -X POST http://localhost:8000/rag/query \
  -H "Content-Type: application/json" \
  -d '{"question": "What was the revenue growth?"}'

Local development

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt
uvicorn app.main:app --reload

Tests

The test suite runs without any API keys — LLM and embeddings are faked:

pytest

Configuration

Set via environment variables or .env (see .env.example):

Variable	Default	Description
ANTHROPIC_API_KEY	–	Anthropic API key (required at runtime)
ANTHROPIC_MODEL	claude-sonnet-4-6	Claude model for all LLM tasks
VOYAGE_API_KEY	–	Voyage AI API key for embeddings (required at runtime)
EMBEDDING_MODEL	voyage-3.5	Voyage embedding model for FAISS
LLM_MAX_TOKENS	8192	Max output tokens per LLM call
CHUNK_SIZE / CHUNK_OVERLAP	1000 / 200	Text splitting for RAG ingestion
VECTOR_STORE_PATH	data/faiss_index	FAISS persistence path (empty = in-memory only)

License

Released under the MIT License.