Memory Packs are portable, immutable contexts for Large Language Models. Too often, data need to be queried at the time of prompt expansion, creating issues od data access, security, and immutability. With Memory Packs, we want to solve this problem by easily creating immutable, portable, and reusable contexts for LLMs. You can think of it as Docker images for LLM's contexts.
Technically, a memory pack is an indexed TAR file that package the context to be used when expanding LLM prompts. The file is a regular tar file but with a special .index file added as the last entry in the tar. The index file list the files inside the tar, one file by line with the byte offset indicating the start of the tar entry where the file is located and the size in bytes of te file content.
Example:
changes.diff:0,256
metadata.json:1024,52
A memory pack contains at least one file: metadata.json. This file is a JSON object containing random properties and is required even if the memory is not defininf any metadata properties. In that case an empty JSON object {} should be put in the metadata file.
The packages in this repository are provinding an implementation of the memory file for javascript and some tooling to build and test memory files.
This is the core library implementing the memory pack and providing an API to programatically create ans use memory packs.
npm install @vertesia/memory
This is a terminal CLI application useful to build and test memory packs.
It builds a memory pack by running a recipe file. The recipe file is a typescript script which may add content to the tar and which must export a JSON object which will be stored in the metadata.json file.
npm install -g @vertesia/memory-cli
The package install an application named memo.
Use memo help to seee the list of commands.
This package must only be imported from the recipe file. It provides a list of the commands to use to build the memory pack.
A recipe is a typescript file that is used to build a memory pack. The recipe can contain any javascript code but it must use the builtin commands to create the memory TAR file.
A memory tpack recipe has the following structure:
import { copy, ... other commands ... } from "@vertesia/memory-commands";
// You can optionally extend an existing memory pack
await from("path/to/memory.tar")
/** then you can build the memory tar content using builtin commands
* You can either use the bult-in commands or `exec` to run a shell command
*/
// ... Your commands here ...
/**
* At the end of the file you must export the memory pack metadata.
* This is a JSON object which holds the properties to be used when creating LLM interactions.
*/
export default {
someProperty: "hello",
...
}The only dependency you need to import in your recipe ts file is the @becmposable/memory-commands package. This package contains all the buil-in commands which are grouped in two categories: execution commands and content loading commands:
- vars - Get the build user variables
- tmpdir - Get or create a temporary directory for use while building
- getBuilder - Get the builder instance.
- from - Extends an existing memory pack (a tar file)
- exec - Execute a shell command.
- copy - Copy a file to the tar. The file content can be transformed using the built-in transformers. See the copy command for more details.
- copyText - Copy an inline text or string variable as a tar entry.
- build - Build an external recipe.
These commands can be used to fetch content from files. The result is returned as a variable which will only fetch the content when used. All these commands supports globs as argument to fretch content from multiple files. When a glob is used the result will alwaysbe an array of variables.
- content - fetch the text content from a text file.
- json - fetch a JSON object from a json file.
- pdf - extract the text content from a PDF file.
- docx - extract the text content from a Docx file.
- media - Load an image with an optional conversion. The image is exported as base64 text, when the variable is referenced in the exported memory metadata.
Signature: vars (): Record<string,any>
A command that returns the user variables which were specified when invoking the memory pack build.
These variables can be used to parametrize the recipe. When using the memo cli application (i.e. @vertesia/memory-cli) you can pass vars to the recipe by using command line parameters which starts with a --var- prefix.
Example: --var-language fr will produce a property language whith a value of 'fr'.
Usage:
import { vars } from "@vertesia/memory-commands"
const { language } = vars();Signature: tmpdir (): string
Get or create a temporary directory. If the directory was not yet created it will be created and its path returned.
If created, the temporary directory will be automatically removed at the end of the build.
Usage:
import { tmpdir } from "@vertesia/memory-commands";
const wd = tmpdir();
await exec(`ls -al > ${wd}/text.txt`);Signature: getBuilder (): Builder
Get the instance of the builder which is used to build the current recipe ts file.
Signature: from (location: string, options?: FromOptions): void
The from command load an existing memory pack tar as a source of the current pack being built. You can filter which files and metadata properties will be included from the soruce tar by specifying a files filter and a metadata properties projection.
The from command is an asynchronous operation so you need to await for its completion.
The shape of the FromOptions is:
{
files?: string[];
projection?: Record<string, boolean | 0 | 1>;
}The files filter is an array of globs as supported by the micromatch library. You can either include files using globs expressions or fiels by prepending the glob expression with an ! character.
The projection is an object which map keys to a truthy or falsy value. You can either use false or true to exclude or include properties not both.
Example:
import { vars, from } from "@vertesia/memory-commands"
await from("./memory-source.tar", {
files: ["images/*.png"],
projection: {
name: true,
language: true,
}
})
// export a new metadata object
export default {
theme: vars.theme || 'dark'
}The memory pack built above will contain all the images/*.png files from the source tar and the metadata JSON object will contain the name and language properties from the source memory pack metadata and will add a new metadata property named theme.
Signature: exec (cmd: string, options?: ExecOptions): Promise<void | string>
Execute a shell command. You can execute a pipe of shell commands or use stdout redirection at the end of the command:
The exec command is asynchronous so you need to use await when incoking it. If no output redirection is specified then the output will be returned as a string from the exec() function
Example:
import { tmpdir, exec } from "@vertesia/memory-commands"
const wd = tmpdir();
const output = await exec("cat some/file | wc -l")
await exec(`cat some/file | wc -l > ${wd}out.txt`)Signature: copy (location: SourceSpec, path: string, options?: CopyOptions): void
This command will copy the content of a file to a tar entry specified by the path argument.
The location argument specify from source from where the content is copied.
It may be either a string or a ContentSource object. ContentSource objects are returned by any of the Content Loading Commands.
If location is a string it will be expected to point to a file path or to a glob expression.
If a glob expression is used the the copy operation will copy the content of all the matching files. You can, in that case, specify a path rewrite expression as the path argument.
Youn can transform the file content when copying it to the tar by using the options argument. The shape of the options is:
export interface CopyOptions {
media?: {
max_hw?: number;
format?: "jpeg" | "png";
};
extractText?: boolean | string;
}You can convert images by specfying a max height or widtth and / or an output image format. For documents like pdf or docx you can specify extractText: "pdf" | "docx" | boolean. If true is specified the type of text extraction will be guessed from the file extension.
Example:
import { exec, copy, tmpdir } from "@vertesia/memory-commands"
const wd = tmpdir();
await exec(`cat some/file > ${wd}/out.txt`)
copy(`${wd}/out.txt`, 'out.txt');
copy('./my-project/src/**/*.ts', './my-project/src!sources/*')The rewrite expression in the example above ./my-project/src!sources/* means: strip the prefix ./my-project/src fromt he copied file and prefix the remaining od the path with the value sources/.
A path rewrite expression is composed of two parts:
- an optional prefix separated by a
!charcater from the rest of the path. If present this prefix will be removed from the matched path. If no prefix is specified then the entrie directory part of the matched path will be removed. - A mandatory suffix which is describing how to rewrite the matched path. The suffix may contain either a wildcard
*which will be replaced with the matched path (without the removed prefix), either a combination of the followinf variables:%n- the file name without the extension%e- the extension%f- the file name including the extension%d- the directory path (not including the removed prefix)%i- the 0 based index of the file in the matched array of files.
Examples:
// copy all .ts files flatened in the sources directory (the directory structure is not preserved)
copy("work/docs/project1/src/**/*.ts", "sources/*")
// copy all .ts files in the sources directory and recreate the subdirectories structure inside src/
copy("work/docs/project1/src/**/*.ts", "work/docs/project1/src!sources/*")
// Remove the work/ prefix and preserver the same subdirectories structure including images/
// and replace the file extension with .png
copy("work/images/**/*.png", "work!%d/%n.jpeg")
// Copy all images inaide a images/ folder without preserving subdirectories and append the index of the image
copy("work/images/**/*.png", "images/%n-%i.%e")In the last example for the mnatched files: work/images/header/home.png and work/images/footer/logo.png the result will be: images/home-0.png, images/logo-1.png
Signature: copyText (text: string, path: string): void
This command will create a new entry in the target memory pack tar using the content you specified through the text argument. The tar entry path is specified by the path argument.
import { exec, copyText } from "@vertesia/memory-commands"
const content = await exec(`cat some/file`)
copyText(content.trim(), 'content.txt')Signature: build (recipePath: string, options?: BuildOptions): void
Build a memory pack from a recipe file. You can use this command if you need to build multiple memory packs.
The BuildOptions options have the following shape:
export interface BuildOptions {
indent?: number;
/**
* the path to save the output. Defaults to 'memory.tar'.
* If no .tar extension is present it will be added
*/
out?: string;
/**
* If set, suppress logs. Defaults to false.
*/
quiet?: boolean;
/**
* If true, compress the output (tar or json) with gzip. Defaults to false.
*/
gzip?: boolean;
/**
* Vars to be injected into the script context as the vars object
*/
vars?: Record<string, any>;
/**
* Optional publish action
* @param file
* @returns the URI of the published memory
*/
publish?: (file: string, name: string) => Promise<string>
}Usage:
import { build, tmpdir } from "@vertesia/memory-commands";
const wd = tmpdir();
await build("./some/recipe.ts", { out: `${wd}}/child-memory`});Signature: content (location: string, encoding?: BufferEncoding): ContentObject | ContentObject[]
If the location is a glob expression then an arry of ContentObject is returned. A ContentObject implements ContentSource so it can be used as the location argument of the copy operation.
The actual cotnent of a file is only fetched (and trasformed) only when the ContentObject is used: either a method of the object is used, either it is assigned to a value of a metadata property.
The content command simply load the content of the given file(s) as a text using an optional encoding. The default encoding is "utf-8".
Signature: json (location: string) => ContentObject | ContentObject[]
Load a JSON object from a json file. When assigned as a metadata property the content will by transformed in a JSON object.
Signature: pdf (location: string) => PdfObject | PdfObject[]
Load a PdfObject form a pdf file. When assigned as a metadata property the PdfObject is transformed to the text representation of the PDF.
Example:
import { pdf } from "@vertesia/memory-commands";
const doc = pdf("./my-doc.pdf")
export default {
textContent: doc
}In the example above the PDF text will be extracted from the pdf and ibjected as the textContent property ion the memory pack metadata.
Load a DocxObject form a docx file. When assigned as a metadata property the DocxObject is transformed to the text representation of the Docx.
Signature: docx (location: string) => DocxObject | DocxObject[]
Signature: json (location: string, options?: MediaOptions) => MediaObject | MediaObject[]
Load a MediaObject form a docx file. When assigned as a metadata property the MediaObject is transformed to text representation of the image which is the base64 encoded image.
The media command accepts an additional argument which has the shape:
export interface MediaOptions {
max_hw?: number;
format?: "jpeg" | "png";
}
You can thus convert the image before using it.
Example:
import { pdf } from "@vertesia/memory-commands";
const images = media("./images/*.jpeg")
export default {
images: images
}In the example above the an array of images encoded as base64 will be injected in the images property pf the memory pack metadata.
To create custom commands you have 2 options. Either you write the command in as a command line application and invoke it through the exec command, either you write a javascript function, package it in an external library and the import it in the recipe ts file.
Here is a template of custom command functions:
import { getBuilder } from "@vertesia/memory-commands";
import fs from "node:fs";
export myCommand(message:string) {
const wd = builder.tmpdir();
// this is the current instance of the builder being used to build the memory pack.
const builder = getBuilder();
// you can add a variable in the build vars:
builder.vars.greeting = 'hello world!';
// or write a file named greeting.txt in the current build tmpdir
fs.writeFileSync(`${wd}/greeting.txt`, "hello world!", "utf-8");
// or create a tar entry named greeting.txt
builder.copyText("hello world!", "greeting.txt");
// refer to Builder API for other operations.
}This example builds a memory pack with the information on which issues were fixed between to commit SHA.
import { copy, copyText, exec, tmpdir, vars } from "@vertesia/memory-commands";
const { start, end } = vars();
if (!start || !end) {
console.error("Please provide start and end tags using --var-start and --var-end");
process.exit(1);
}
const wd = tmpdir();
console.log(`Retrieving issues between ${start}and ${end}...`)
// Get list of commit logs containing '#' between the two tags and extract unique issue numbers
const issue_numbers = await exec(`git log ${start}..${end} --oneline | grep -o '#[0-9]\\+' | sed 's/#//' | sort -u`) as string;
for (const issue of issue_numbers.trim().split("\n")) {
console.log(`Processing issue #${issue}`)
const issue_content = await exec(`gh issue view ${issue}`) as string;
copyText(issue_content, `issues/${issue}.txt`);
}
console.log("Generating diff");
await exec(`git diff --submodule=diff ${start}...${end} > ${wd}/range_diff.txt`)
copy(`${wd}/range_diff.txt`, "range_diff.txt");
export default {
from_version: start,
release_version: end,
}When using a memory pack you will likely want to select the metadata object and some files inside the tar. At each run you may need to access different content files. The content you want to access will be exported as a JSON object by providing a content mapping.
TODO