Tencent/WeKnora
WeKnora is an LLM-powered modular framework for deep document understanding, semantic retrieval, and context-aware question answering using the Retrieval-Augmented Generation paradigm.
Awesome AI Agents › Document Analytics Platforms
johnsonhk88/AI-Bank-Statement-Document-Automation-By-LLM-And-Personal-Finanical-Analysis-Prediction
This project automates the handling of bank statements end to end, from raw PDF to answered questions about personal finances. Documents are parsed with a combination of YOLO layout detection, OCR and LLM-based table extraction through PyMuPDF, which turns statement pages into structured transaction records. The extracted data then feeds three parallel agent harnesses that the author uses to compare approaches: a CrewAI baseline with domain skills, LangChain Deep Agents built on LangGraph, and Hermes running inside a Docker sandbox. The skills are organised by task, covering bank-statement parsing, personally identifiable information redaction, financial analysis, RAG query handling and output formatting. Redaction runs before anything is embedded, so the vector store, either Qdrant or Chroma, never receives raw account identifiers. On top of the extraction layer the system categorises income and expenses, performs trend analysis and prediction, and answers natural language questions about the statement history. The runtime is a conventional production stack rather than a notebook demo: FastAPI with async SQLAlchemy, Alembic migrations and Pydantic v2 for the API, PostgreSQL for storage, Celery and Redis for asynchronous document processing, JWT authentication, and a React, TypeScript and Vite single-page frontend. Model access goes through LiteLLM, which lets the same code run against local models in LM Studio or Ollama as well as hosted providers such as OpenAI, DeepSeek and OpenRouter, and MLflow traces the CrewAI path. Infrastructure files cover Docker Compose and Kubernetes, with optional NVIDIA GPU support for embeddings and local inference. The README documents practical limits for local models, recommending at least nine billion parameters and sixteen thousand tokens of context, and warning that small local models often fail strict JSON output. A Jupyter notebook is included for CrewAI experimentation.
Also in Document Analytics Platforms
StarTrail-org/PixelRAG
Visual retrieval-augmented generation pipeline that renders web pages, PDFs and images to screenshot tiles and retrieves over them with a LoRA-tuned Qwen3-VL embedding model.
georgia-tech-db/evadb
EvaDB is an open-source database system that empowers developers to build AI-powered applications by integrating advanced AI models and optimizations within a SQL-based querying framework for both structured and unstructured data.
loggie-io/loggie
Loggie is a lightweight, cloud-native data transfer agent and aggregator designed for scalable, high-performance log collection, processing, and monitoring in Kubernetes environments.
netboxlabs/pktvisor
pktvisor is a dynamic network observability agent that analyzes high volume network traffic and generates opentelemetry metrics for real-time and centralized monitoring.
genomoncology/FuzzTypes
FuzzTypes is a Pydantic extension that provides autocorrecting annotation types for enhanced data validation and normalization, enabling automatic correction and structured data composition in Python applications.