Skip to content

DEVELOPER.md

Latest commit

 

History

History
89 lines (61 loc) · 3.71 KB

DEVELOPER.md

File metadata and controls

89 lines (61 loc) · 3.71 KB
Raw

Developer notes

These are some notes documenting annoying or complex parts of developing for ReVa and the general architecture of ReVa. It is assumed you read the README.md before reading this.

Design

ReVa is three components:

  • An extension/plugin (Ghidra, Binary Ninja, etc.)
  • reva-server - The LLM interface
  • reva-chat - The chat UI

These three components talk to each other using gRPC. The extension launches reva-server when it starts and manages its lifetime. reva-chat communicates with the extension and reva-server.

There is a one to one relationship between the extension and reva-server instances. There is a many to one relationship of reva-chat to the extension and server.

reva-server is written in Python and can use Ollama and OpenAI for inference. The Ollama server can be local or remote.

The protocol directory contains the protobuf definitions for each message type and a definition for the functions provided. Services are small and can be hosted on either the extension or the inference side. This allows the LLM to ask the extension for information, or the extension to ask the LLM for analysis.

The reva-chat client first locates the running extension, and then gets details of the reva-server started by that extension. The /tmp/.reva (or, on Windows, the %TEMP%/.reva) directory is used to find running extensions.

Building the Ghidra extension

You will need:

Note: This means Gradle 7 for Ghidra and gRPC support!

The Ghidra extension is built like any other Ghidra extension. Running gradle will generate a dist/ directory and you can install the extension from there.

If you are running a UNIX system, there is a helper script gext-build that will build and install the extestion (replacing older versions) in one step.

reva-server -> Extension

These tools allow ReVa to interact with an RE tool and the binary/project it has open.

These include:

  • Getting decompilation for a function
  • Renamaing variables or symbols
  • Getting a list of functions

The implementation for these are in the Ghidra extension's Handlers directory.

The client code is in the reva-server's api_server_tools.

Extension -> reva-server

These tools allow the RE tool to make queries to the LLM.

These include:

  • Asking for a better variable name
  • Explaining some code that is highlighted
  • Setting a comment

These typically just forward some context information to ReVa, and ReVa will format a prompt and send it to the LLM. Usually the LLM will then use ReVa -> Extension tools to perform any actions required.

These are implemented in reva-server.

The client code is mostly in the UI code of the Ghidra extenstion.

Adding a new Tool

We use the gRPC library to talk between the RE Tool and ReVa.

https://grpc.io/docs/languages/java/quickstart/

  1. Add a protocol entry to the protocol directory.
  2. Add a server implementation to the correct location
  3. For a reva-server -> Extension tool, Handlers in the Ghidra extension
  4. For an Extension -> reva-server tool, api_server_tools
  5. Add a stub to call the server.