Gemini doesn't work well when trying to output json objects. Using it
to output raw json strings with complex, multi-line structures
requires more intense clean-up of raw json string for parsing
Collect, display and store running costs & accuracy of eval run.
This provides more insight into eval runs during execution instead of
having to wait until the eval run completes.
- Previously online chat actors, director tests only worked with openai.
This change allows running them for any supported onlnie provider
including Google, Anthropic and Openai.
- Enable online/offline chat actor, director in two ways:
1. Explicitly setting KHOJ_TEST_CHAT_PROVIDER environment variable to
google, anthropic, openai, offline
2. Implicitly by the first API key found from openai, google or anthropic.
- Default offline chat provider to use Llama 3.1 3B for faster, lower
compute test runs
- Set output mode to single string. Specify output schema in prompt
- Both thesee should encourage model to select only 1 output mode
instead of encouraging it in prompt too many times
- Output schema should also improve schema following in general
- Standardize variable, func name of io selector for readability
- Fix chat actors to test the io selector chat actor
- Make chat actor return sources, output separately for better
disambiguation, at least during tests, for now
- Evaluate khoj on random 200 questions from each of google frames and openai simpleqa benchmarks across *general*, *default* and *research* modes
- Run eval with Gemini 1.5 Flash as test giver and Gemini 1.5 Pro as test evaluator models
- Trigger eval workflow on release or manually
- Make dataset, khoj mode and sample size configurable when triggered via manual workflow
- Enable Web search, webpage read tools during evaluation
- Just load the raw csv from OpenAI bucket. Normalize it into FRAMES format
- Improve docstring for frames datasets as well
- Log the load dataset perf timer at info level
- Default to evaluation decision of None when either agent or
evaluator llm fails. This fixes accuracy calculations on errors
- Fix showing color for decision True
- Enable arg flags to specify output results file paths
Previously the batch start index wasn't being passed so all batches
started in parallel were showing the same processing example index
This change doesn't impact the evaluation itself, just the index shown
of the example currently being evaluated
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.
Overview
---
- Put context into separate user message before sending to chat model.
This should improve model response quality and truncation logic in code
- Pass online context from chat history to chat model for response.
This should improve response speed when previous online context can be reused
- Improve format of notes, online context passed to chat models in prompt.
This should improve model response quality
Details
---
The document, online search context are now passed as separate user
messages to chat model, instead of being added to the final user message.
This will improve
- Models ability to differentiate data from user query.
That should improve response quality and reduce prompt injection
probability
- Make truncation logic simpler and more robust
When context window hit, can simply pop messages to auto truncate
context in order of context, user, assistant message for each
conversation turn in history until reach current user query
The complex, brittle logic to extract user query from context in
last user message isn't required.
* Create explicit flow to enable the free trial
The current design is confusing. It obfuscates the fact that the user is on a free trial. This design will make the opt-in explicit and more intuitive.
* Use the Subscription Type enum instead of hardcoded strings everywhere
* Use length of free trial in the frontend code as well
Given the LLM landscape is rapidly changing, providing a good default
set of options should help reduce decision fatigue to get started
Improve initialization flow during first run
- Set Google, Anthropic Chat models too
Previously only Offline, Openai chat models could be set during init
- Add multiple chat models for each LLM provider
Interactively set a comma separated list of models for each provider
- Auto add default chat models for each provider in non-interactive
model if the {OPENAI,GEMINI,ANTHROPIC}_API_KEY env var is set
- Do not ask for max_tokens, tokenizer for offline models during
initialization. Use better defaults inferred in code instead
- Explicitly set default chat model to use
If unset, it implicitly defaults to using the first chat model.
Make it explicit to reduce this confusion
Resolves#882
# Summary of Changes
* New UI to show preview of image uploads
* ChatML message changes to support gpt-4o vision based responses on images
* AWS S3 image uploads for persistent image context in conversations
* Database changes to have `vision_enabled` option in server admin panel while configuring models
* Render previously uploaded images in the chat history, show uploaded images for pending msgs
* Pass the uploaded_image_url through to subqueries
* Allow image to render upon first message from the homepage
* Add rendering support for images to shared chat as well
* Fix some UI/functionality bugs in the share page
* Convert user attached images for chat to webp format before upload
* Use placeholder to attached image for data source, response mode actors
* Update all clients to call /api/chat as a POST instead of GET request
* Fix copying chat messages with images to clipboard
TLDR; Add vision support for openai models on Khoj via the web UI!
---------
Co-authored-by: sabaimran <narmiabas@gmail.com>
Co-authored-by: Debanjum Singh Solanky <debanjum@gmail.com>
GPT-4o-mini is cheaper, smarter and can hold more context than
GPT-3.5-turbo. In production, we also default to gpt-4o-mini, so makes
sense to upgrade defaults and tests to work with it
- Major
- Improve doc search actor performance on vague, random or meta questions
- Pass user's name to document and online search actors prompts
- Minor
- Fix and improve openai chat actor tests
- Remove unused max tokns arg to extract qs func of doc search actor
### Overview
Support exclude file filter in user search queries
### Details
- All of the exclude file filter terms need to be satisfied
- Any one of the include file filter terms should be satisfied
### Example
- **Search Query**: *what happened yesterday? -file:"tasks.org" -file:"work.md" file:"diary.org" file:"journal.org*
- **Behavior**: Query will try find relevant notes in any of `journal.org` or `diary.org` and not in `tasks.org` and not in `work.md`
### Details
* Add support for exclusion file filters
* Translate file filter to valid Django DB entry filter regex
* Exclude all files when multiple exclude file filter in query
Previously we were applying an "Or" filter, which would exclude any
file mentioned in a query with multiple exclude file filter.
This is not what we naturally mean when we ask excluding a file in a query
* Rename, rearrange, deduplicate and add file filter tests
Closes#728
---------
Co-authored-by: Debanjum Singh Solanky <debanjum@gmail.com>
- Use new form of passing doc references to now passing chat actor
test
- Fix message list generation from conversation logs provided
Strangely the parent conversation_log gets passed down to
message_to_log func when the kwarg is not explicitly specified
