- Why
We need better, automated evals to measure performance shifts of Khoj
across prompt, model and capability changes.
Google's FRAMES benchmark evaluates multi-step retrieval and reasoning
capabilities of AI agents. It's a good starter benchmark to evaluate Khoj.
- Details
This PR adds an eval script to evaluate Khoj responses on the the FRAMES
benchmark prompts against the ground truth provided by it.
Script allows configuring sample size, batch size, sampling queries from the
eval dataset.
Gemini is used as an LLM Judge to auto grade Khoj responses vs ground truth
data from the benchmark.
This was previously required, but now it's only usefuly for more
advanced settings, not typical for self-hosting users.
With recent updates, the user's selected chat model is used for both
Khoj's train of thought and response. This makes it easy to
switch your preferred chat model directly from the user settings
page and not have to update this in the admin panel as well.
Reflect these code changse in the docs, by removing the unnecessary
step for self-hosted users to create a server chat setting when using
an OpenAI proxy service like Ollama, LiteLLM etc.
Now that models have much larger context windows, we can reasonably include full texts of certain files in the messages. Do this when an explicit file filter is set in a conversation. Do so in a separate user message in order to mitigate any confusion in the operation.
Pipe the relevant attached_files context through all methods calling into models.
We'll want to limit the file sizes for which this is used and provide more helpful UI indicators that this sort of behavior is taking place.
- The server has moved to a model of standardization for the embeddings generation workflow. Remove references to the support for differentiated models.
- The migration script fo ra new model needs to be updated to accommodate full regeneration.
Google's FRAMES benchmark evaluates multi-step retrieval and reasoning
capabilities of an agent.
The script uses Gemini as an LLM Judge to evaluate Khoj responses to
the FRAMES benchmark prompts against the ground truth provided by it.
- Improve chat actors and their prompts for research mode.
- Add documentation to enable the code tool when self-hosting Khoj
- Edit Chat Messages
- Store Turn Id in each chat message.
- Expose API to delete chat message.
- Expose delete chat message button to turn delete chat message from web app
- Set LLM Generation Seed for Reproducible Debugging and Testing
- Setting seed for LLM generation is supported by Llama.cpp and OpenAI models.
This can (somewhat) restrain LLM output
- Getting fixed responses for fixed inputs helps test, debug longer reasoning chains like used in advanced reasoning
## Overview
Khoj can now go into research mode and use a python code interpreter. These are experimental features that are being released early for feedback and testing.
- Research mode allows Khoj to dynamically select the tools it needs to best answer the question. It is also allowed more iterations to get to a satisfactory answer. Its more dynamic train of thought is shown to improve visibility into its thinking.
- Adding ability for Khoj to use a python code interpreter is an adjacent capability. It can help Khoj do some data analysis and generate charts for you. A sandboxed python to run code is provided using [cohere-terrarium](https://github.com/cohere-ai/cohere-terrarium?tab=readme-ov-file), [pyodide](https://pyodide.org/).
## Analysis
Research mode (significantly?) improves Khoj's information retrieval for more complex queries requiring multi-step lookups but takes longer to run. It can research for longer, requiring less back-n-forth with the user to find an answer.
Research mode gives most gains when used with more advanced chat models (like o1, 4o, new claude sonnet and gemini-pro-002). Smaller models improve their response quality but tend to get into repetitive loops more often.
## Next Steps
- Get community feedback on research mode. What works, what fails, what is confusing, what'd be cool to have.
- Tune Khoj's capabilities for longer autonomous runs and to generalize across a larger range of model sizes
## Miscellaneous Improvements
- Khoj's train of thought is saved and shown for all messages, not just the latest one
- Render charts generated by Khoj and code running using the code tool on the web app
- Align chat input color to currently selected agent color
- Dedent code for readability
- Use better name for in research mode check
- Continue to remove inferred summarize command when multiple files in
file filter even when not in research mode
- Continue to show select information source train of thought.
It was removed by mistake earlier
