Skip to content

Latest commit

 

History

History
71 lines (47 loc) · 3.7 KB

chainlit.md

File metadata and controls

71 lines (47 loc) · 3.7 KB

Introduction to Jar3d

What is Jar3d?

Jar3d is a versatile agent for research intensive tasks that combines chain-of-reasoning, Meta-Prompting, and Agentic RAG techniques.

  • It features integrations with popular providers and open-source models, allowing for 100% local operation given sufficient hardware resources.

How to use

Jar3d will introduce itself and help you refine your requirements. Once you have delivered all the required type /end. Jar3d will pass your requirements on to the MetaExpert who will work with its team of experts to deliver on your requirements.

Use Cases & Applications

  • Long-running research tasks, writing literature reviews, newsletters, etc.
  • Potential adaptation for use with internal company documents, requiring no internet access.
  • Can function as a research assistant or a local version of services like Perplexity.
  • Market research
  • Sourcing goods (i.e. Find me the chepaest A5 Wagyu beef from French suppliers)

For setup instructions, please refer to the Setup Jar3d guide.

How Jar3d works

Prompt Engineering

Jar3d utilizes two powerful prompts written entirely in Markdown:

  1. Jar3d Meta-Prompt
  2. Jar3d Requirements Prompt

Both prompts incorporate adaptations of the Chain of Reasoning technique.

Jar3d Architecture

The Jar3d architecture incorporates aspects of Meta-Prompting, Agentic RAG, and an adaptation of Chain of Reasoning.

Jar3d's Retrieval Mechanism for Internet Research

This system employs a sophisticated retrieval mechanism for conducting internet research. The process involves several steps, utilizing various tools and techniques to ensure comprehensive and relevant results.

1. Web Page Discovery

  • Utilizes the SERPER tool to find relevant web pages.
  • Employs an LLM-executed search algorithm, expressed in natural language.
  • Each iteration of the algorithm generates a search query for SERPER.
  • SERPER returns a search engine results page (SERP).
  • Another LLM call selects the most appropriate URL from the SERP.
  • This process is repeated a predetermined number of times to compile a list of URLs for in-depth research.

2. Content Extraction and Chunking

  • Employs LLM Sherpa as a document ingestor.
  • Intelligently chunks the content from each URL in the compiled list.
  • Results in a corpus of chunked text across all accumulated URLs.

3. Text Embedding

  • Embeds the chunked text using a locally hosted model from FastEmbed.
  • Indexes embeddings in an in-memory FAISS vector store.

4. Similarity Search

  • Performs retrieval using a similarity search over the FAISS vector store.
  • Utilizes cosine similarity between indexed embeddings and the meta-prompt (written by the meta-agent).
  • Retrieves the most relevant information based on this similarity measure.

5. Re-ranking

  • Leverages FlashRank as a locally hosted re-ranking service.
  • FlashRank uses cross-encoders for more accurate assessment of document relevance to the query.

6. Final Selection

  • Selects a designated percentile of the highest-scoring documents from the re-ranked results.
  • Passes this final set of retrieved documents to the meta-agent for further processing or analysis.