ELISA

ELISA (Emacs Lisp Information System Assistant) is a system designed to provide informative answers to user queries by leveraging a Retrieval Augmented Generation (RAG) approach.

Data Sources and Processing

ELISA can access and process information from multiple sources, including:

• Local Files: ELISA can analyze text content within local files, enabling it to retrieve information specific to a user’s projects or documents.
• Info manuals: ELISA has access to the comprehensive Emacs info manuals covering Emacs itself, Emacs Lisp, and various Emacs Lisp packages.
• Web Search: ELISA integrates web search capabilities to provide access to a vast pool of publicly available information.

RAG Methodology

ELISA implements a RAG framework to process and respond to queries. This involves:

1. Data Parsing: Input data is parsed and organized into structured collections for efficient retrieval.
2. Contextual Analysis: When a query is received, ELISA analyzes the context within relevant data collections to identify passages containing potentially useful information.
3. Response Generation: ELISA synthesizes a response based on the identified contextual quotes, aiming to provide a comprehensive and accurate answer to the user’s question.

Collections

ELISA operates on collections. Collection has name and contains quotes from parsed documents. Every web search, every directory parsing creates new collection. ELISA searches in enabled collection, adds relevant information to context and let LLM answer user query.

PDF files support

For pdf, doc and other complex documents support used Apache Tika.

Installation

You need emacs 29.2 or newer to use this package.

This package now on MELPA and you can just M-x package-install elisa.

System dependencies

Sqlite extensions

Then you need to download sqlite-vss. You can do it manually from https://github.com/asg017/sqlite-vss/releases or by calling M-x elisa-download-sqlite-vss.

Large language models

You can use this package with different llm providers. By default it uses ollama provider both for embeddings and chat. If you ok with it, you need to install ollama and pull used models:

ollama pull nomic-embed-text
ollama pull sskostyaev/openchat:8k-rag

Second model is just openchat with exactly 2 tweaks: context window extended to 8k and temperature set to 0 to better usage for RAG (Retrieval Augmented Generation). You can try other models, for example:

• all-minilm for embeddings
• sskostyaev/mistral:32k-rag for chat
• other models or providers
• create your own model

I prefer this models:

ollama pull gemma2:9b-instruct-q6_K
ollama pull qwen2.5:3b
ollama pull chatfire/bge-m3:q8_0

Complex documents

For pdf, doc etc. you need to run Apache Tika service locally. You can do it using docker:

docker run -d -p 127.0.0.1:9998:9998 apache/tika:latest-full

Reranker

Reranker disabled by default to decrease number of system dependencies, but it improves quality of retrieving and answers significantly. You can find installation instructions here. Recommended.

Web search provider

By defauld duckduckgo used for web search. But I prefer searxng. The simplest way to use searxng is docker. You need to enable json format in settings.

Parse info manuals

Create index for builtin, external or all info manuals by one of this commands:

• elisa-async-parse-builtin-manuals
• elisa-async-parse-external-manuals
• elisa-async-parse-all-manuals

This can take some time.

Commands

elisa-chat

Entrypoint. Makes hybrid search in enabled collections, add founded quotes into context and query llm for prompt. Uses ellama under the hood.

elisa-download-sqlite-vss

Download sqlite vss extension to provide similarity search.

elisa-async-parse-builtin-manuals

Parse builtin emacs info manuals asyncronously. Can take long time.

elisa-async-parse-external-manuals

Parse external emacs info manuals asyncronously. Can take long time.

elisa-async-parse-all-manuals

Parse all emacs info manuals asyncronously.

One of parse functions should be called before elisa-chat to create index.

elisa-web-search

Search the web and answer to user query based on found information.

How it works

Search the web for user query. Create new collection with user query as name. Parse web pages to this new collection. Search in this collection. Add related information to context. Ask llm to answer user query based on provided context.

elisa-async-parse-directory

Parse directory as new collection. Can take long time. Works asyncronously and incrementally.

elisa-reparse-current-collection

Incrementally reparse current directory collection. It does nothing if buffer file not inside one of existing collections. Works asyncronously.

elisa-create-empty-collection

Create new empty collection.

elisa-add-file-to-collection

Add file to collection.

elisa-add-webpage-to-collection

Add webpage to collection.

elisa-enable-collection

Enable collection for elisa-chat.

elisa-disable-collection

Disable collection.

elisa-disble-all-collections

Disable all collections.

elisa-remove-collection

Removes collection and all its data from index.

Configuration

Example configuration.

(use-package elisa
  :init
  (setopt elisa-limit 5)
  ;; reranker increases answer quality significantly
  (setopt elisa-reranker-enabled t)
  ;; prompt rewriting may increase quality of answers
  ;; disable it if you want direct control over prompt
  (setopt elisa-prompt-rewriting-enabled t)
  (require 'llm-ollama)
  ;; gemma 2 works very good in my use cases
  ;; it also boasts strong multilingual capabilities
  ;; (setopt elisa-chat-provider
  ;; 	  (make-llm-ollama
  ;; 	   :chat-model "gemma2:9b-instruct-q6_K"
  ;; 	   :embedding-model "chatfire/bge-m3:q8_0"
  ;; 	   ;; set context window to 8k
  ;; 	   :default-chat-non-standard-params '(("num_ctx" . 8192))))
  ;;
  ;; qwen 2.5 3b works good in my test cases and provide longer context
  (setopt elisa-chat-provider
	  (make-llm-ollama
	   :chat-model "qwen2.5:3b"
	   :embedding-model "chatfire/bge-m3:q8_0"
	   :default-chat-temperature 0.1
	   :default-chat-non-standard-params '(("num_ctx" . 32768))))
  ;; this embedding model has stong multilingual capabilities
  (setopt elisa-embeddings-provider (make-llm-ollama :embedding-model "chatfire/bge-m3:q8_0"))
  :config
  ;; searxng works better than duckduckgo in my tests
  (setopt elisa-web-search-function 'elisa-search-searxng))

ELISA Custom Variables

General Settings

• elisa-embeddings-provider:
  • Description: LLM provider to generate embeddings for text.
  • Default: (make-llm-ollama :embedding-model "nomic-embed-text")
• elisa-chat-provider:
  • Description: LLM provider used for chat interactions.
  • Default: (make-llm-ollama :chat-model "sskostyaev/openchat:8k-rag" :embedding-model "nomic-embed-text")
• elisa-db-directory:
  • Type: Directory
  • Description: Specifies the directory where ELISA stores its database.
  • Default: (file-name-concat user-emacs-directory "elisa") (within your Emacs config directory)
• elisa-limit:
  • Type: Integer
  • Description: Controls the number of quotes passed to the LLM context for generating an answer.
  • Default: 5
• elisa-find-executable:
  • Type: String
  • Description: Path to the find command executable. Used for locating files.
  • Default: “find”

File System and Database Management

• elisa-tar-executable:
  • Type: String
  • Description: Path to the tar command executable. Used for archiving files.
  • Default: “tar”
• elisa-sqlite-vss-version:
  • Type: String
  • Description: Version of the SQLite VSS extension.
• elisa-sqlite-vss-path:
  • Type: File path
  • Description: Path to the SQLite VSS extension file.
• elisa-sqlite-vector-path:
  • Type: File path
  • Description: Path to the SQLite Vector extension file.

Text Processing and Semantic Splitting

• elisa-semantic-split-function:
  • Type: Function
  • Description: Function used to split text into semantically meaningful chunks.
  • Default: elisa-split-by-paragraph
• elisa-prompt-rewriting-enabled:
  • Type: Boolean
  • Description: Enables or disables prompt rewriting for better retrieving.
  • Default: t (enabled)
• elisa-chat-prompt-template:
  • Type: String
  • Description: Template used for constructing the chat prompt.
• elisa-rewrite-prompt-template:
  • Type: String
  • Description: Template used for rewriting prompts for better retrieval.

Web Search and Integration

• elisa-searxng-url:
  • Type: String
  • Description: URL of your SearXNG instance.
  • Default: “http://localhost:8080/”
• elisa-pandoc-executable:
  • Type: String
  • Description: Path to the pandoc command executable. Used for converting documents to text.
  • Default: “pandoc”
• elisa-webpage-extraction-function:
  • Type: Function
  • Description: Function used to extract the content from a webpage.
  • Default: elisa-get-webpage-buffer
• elisa-web-search-function:
  • Type: Function
  • Description: Function responsible for performing web searches using the provided prompt.
  • Default: elisa-search-duckduckgo
• elisa-web-pages-limit:
  • Type: Integer
  • Description: Maximum number of web pages to parse during a search.
  • Default: 10

Reranking

• elisa-breakpoint-threshold-amount:
  • Type: Float
  • Description: Threshold used for controlling the granularity of semantic splitting.
  • Default: 0.4
• elisa-reranker-enabled:
  • Type: Boolean
  • Description: Enables or disables reranking, which can improve retrieval quality by ranking retrieved quotes based on relevance.
  • Default: nil (not set)
• elisa-reranker-url:
  • Type: String
  • Description: URL of the reranking service.
  • Default: “http://127.0.0.1:8787/”
• elisa-reranker-similarity-threshold:
  • Type: Float
  • Description: Similarity threshold for reranking. Quotes below this threshold will be filtered out. If not set all ellama-limit quotes will be added to context.
  • Default: 0
• elisa-reranker-limit:
  • Type: Integer
  • Description: Number of quotes to send to the reranker.
  • Default: 20

File Parsing and Exclusion

• elisa-ignore-patterns-files:
  • Type: List of strings
  • Description: List of file name patterns (e.g., .gitignore) used to ignore files during parsing.
  • Default: (".gitignore" ".ignore" ".rgignore")
• elisa-ignore-invisible-files:
  • Type: Boolean
  • Description: Toggles whether invisible files and directories should be ignored during parsing.
  • Default: t (true)
• elisa-tika-url:
  • Type: String
  • Description: Apache tika url for file parsing.
  • Default: "http://localhost:9998/"
• elisa-complex-file-extraction-function:
  • Type: Function
  • Description: Function to get buffer with complex file (like pdf, odt etc.) content.
  • Default: #'elisa-parse-with-tika-buffer
• elisa-supported-complex-document-extensions:
  • Type: List of strings
  • Description: Supported complex document file extensions.
  • Default: '("doc" "dot" "ppt" "xls" "rtf" "docx" "pptx" "xlsx" "xlsm" "pdf" "epub" "msg" "odt" "odp" "ods" "odg" "docm")

ELISA Chat Collections

• elisa-enabled-collections:
  • Type: List of strings
  • Description: Specifies which collections are enabled for chat interactions.
  • Default: ("builtin manuals" "external manuals")

Contributions

To contribute, submit a pull request or report a bug. This library is planned to be part of GNU ELPA; major contributions must be from someone with FSF papers. Alternatively, you can write a module and share it on a different archive like MELPA.