RAG (Retrieval Augmented Generation) is a method to combine a search
engine with a LLM (Large Language Model). When a new prompt is
submitted, a search engine injects knowledge from a search into the
content. This is done using a reverse proxy between the Chat Client and
the LLM. In this case, we used the following software:
LLM Backend - Ollama:
https://github.com/ollama/ollama
Install ollama and then load two required LLM models
with the following commands:
ollama pull phi3:3.8b
ollama pull llama3:8b
Chat Client - susi_chat:
https://github.com/susiai/susi_chat
just clone the repository and the open the file
susi_chat/chat_terminal/index.html
in your browser. This displays a chat terminal.
In this terminal, run the following command:
host http://localhost:8090
This sets the LLM backend to your YaCy peer.
Then start YaCy. It will provide the LLM endpoint to the client
while using ollama in the backend. It then injects search results
only from the local Solr index, not from the p2p network (so far).
The export file can now be many files, called chunks.
By default still only one chunk is exported.
This function is required in case that the exported files shall be
imported to an elasticsearch/opensearch index. The bulk import function
of elasticsearch/opensearch is limited to 100MB. To make it possible to
import YaCy files, those must be splitted into chunks. Right now we
cannot estimate the chunk size as bytes, only as number of documents.
The user must do experiments to find out the optimum chunk max size,
like 50000 docs per chunk. Try this as first attempt.
You can now import zim files into YaCy by simply moving them
to the DATA/SURROGATE/IN folder. They will be fetched and after
parsing moved to DATA/SURROGATE/OUT.
There are exceptions where the parser is not able to identify the
original URL of the documents in the zim file. In that case the file
is simply ignored.
This commit also carries an important fix to the pdf parser and an
increase of the maximum parsing speed to 60000 PPM which should make it
possible to index up to 1000 files in one second.
Michael Peter Christen