* Add checksums to verify the correct model is downloaded as expected
- This should help debug issues related to corrupted model download
- If download fails, let the application continue
* If the model is not download as expected, add some indicators in the settings UI
* Add exc_info to error log if/when download fails for llamav2 model
* Simplify checksum checking logic, update key name in model state for web client
# Incoming
## Major
### Fix Prompt Size Exceeded Issue
- Fix issues related to prompt size, Closes#386. Use the correct tokenizer to calculate whether the input needs to be truncated or not.
### Improve Llama 2 Model Download
- Use the correct download link for LlamaV2 -- should have been using the small model, but was using the medium
- Add better downloading logic to retry download if it failed, Closes#379
### Fix Segmentation Fault due to Race
- Add a lock around generating chat responses from the offline model to avoid segmentation faults. Closes#367.
- Add a loading symbol to the web chat UI when the model is thinking. Closes#392
### Improve Chat Response Latency
- Improve performance of offline chat by increasing batch size (via `n_batch`) to automatically engage more cores/GPU, using smaller model and fixing prompt vs response token generation numbers. Closes#363
### Fix Fake Dialogue Continuation
- Fix formatting of user query with offline chat, this was contributing to #398
- Stop Llama 2 from Creating Fake Dialogue Continuations. Closes#398
## Minor
- Improve default message for Chat window on web when it's not configured. Include hint to use offline chat.
- Add null check in `perform_chat_checks` method
- Add offline chat director unit tests
## Performance Analysis (Time to First Token)
| | v0.10.0 | this branch |
|-|-|-|
| Query 1 | 52s | 28s |
| Query 2 | 33s| 42s |
| Query 3 | 67s| 38s|
It would previously some times start generating fake dialogue with
it's internal prompt patterns of <s>[INST] in responses.
This is a jarring experience. Stop generation response when hit <s>
Resolves#398
- Use same batch_size in extract question actor as the chat actor
- Log final location the chat model is to be stored in, instead of
it's temp filename while it is being downloaded
- Fix download url -- was mapping to q3_K_M, but fixed to use q4_K_S
- Use a proper Llama Tokenizer for counting tokens for truncation with Llama
- Add additional null checks when running
Previously the system message was getting dropped when the context
size with chat history would be more than the max prompt size
supported by the cat model
Now only the previous chat messages are dropped or the current
message is truncated but the system message is kept to provide
guidance to the chat model
* Add support for configuring/using offline chat from within Obsidian
* Fix type checking for search type
* If Github is not configured, /update call should fail
* Fix regenerate tests same as the update ones
* Update help text for offline chat in obsidian
* Update relevant description for Khoj settings in Obsidian
* Simplify configuration logic and use smarter defaults
- Configure using Offline Chat from Emacs:
- Enable, Disable Offline Chat from Emacs
- Use: Enable offline chat with `(setq khoj-chat-offline t)' during khoj setup
- Benefits: Offline chat models are better for privacy but not great at answering questions
* Let Offline chat override OpenAI API settings
* Download the offline model whenever offline chat is enabled
* Add progressbar for download for llamav2 model to track progress
* Change ordering of n due to switch of default processor
* Flip ordering of offline/openai checks when extracting questions from query
* Working example with LlamaV2 running locally on my machine
- Download from huggingface
- Plug in to GPT4All
- Update prompts to fit the llama format
* Add appropriate prompts for extracting questions based on a query based on llama format
* Rename Falcon to Llama and make some improvements to the extract_questions flow
* Do further tuning to extract question prompts and unit tests
* Disable extracting questions dynamically from Llama, as results are still unreliable
* Add support for gpt4all's falcon model as an additional conversation processor
- Update the UI pages to allow the user to point to the new endpoints for GPT
- Update the internal schemas to support both GPT4 models and OpenAI
- Add unit tests benchmarking some of the Falcon performance
* Add exc_info to include stack trace in error logs for text processors
* Pull shared functions into utils.py to be used across gpt4 and gpt
* Add migration for new processor conversation schema
* Skip GPT4All actor tests due to typing issues
* Fix Obsidian processor configuration in auto-configure flow
* Rename enable_local_llm to enable_offline_chat
* Add docs for more organized, accessible information detailing Khoj setup
* Delete duplicated files
* Add a coverpage without enabling it. Add logo and theme
* Remove obsidian README.md
* Add plausible script to index.html via docsify
## Stabilize and Simplify Content Indexing
### Major Updates
- 9bcca43 Unify logic to update entries when indexing from scratch or incrementally
- 89c7819 Unify logic to update embeddings when indexing from scratch or incrementally
- 6a0297c Stable sort new entries when marking entries for update
- 58d86d7 Unify logic to configure server from API or on server start
- Create tests to ensure old entries, embeddings in index are unaffected on adding new entries
- Refer: 1482fd4, 7669b85, 88d1a29
- ad41ef3 Make normalization of embeddings configurable to test this in c73feeb
### Minor Updates
- 1673bb5 Add todo state to compiled form of each entry
- 6e70b91 Remove unused `dump_jsonl` helper method
- 7ad9603 Improve naming of lock
- b02323a Improve naming text search test methods
Resolves#190
Previous regenerate mechanism did not deduplicate entries with same key
So entries looked different between regenerate and update
Having single func, mark_entries_for_update, to handle both scenarios
will avoid this divergence
Update all text_to_jsonl methods to use the above method for
generating index from scratch
Reuse Search Models across Content Types to reduce Memory Consumption
- Memory consumption now only scales with search models used, not with content types.
Previously each content type had it's own copy of the search ML models.
That'd result in 300+ Mb per enabled text content type
- Split model state into 2 separate state objects, `search_models` and `content_index`.
This allows loading text_search and image_search models first
and then reusing them across all content_types in content_index
- The change should cut down memory utilization quite a bit for most users.
I see a >50% drop in memory utilization on my Khoj instance.
But this will vary for each user based on the amount of content indexed vs number of plugins enabled.
- This change does not solve the RAM utilization scaling with size of the index,
as the whole content index is still kept in RAM while Khoj is running
Should help with #195, #301 and #303
Wrap acquire/release locks in try/catch/finally when updating content
index and search models to prevent lock not being released on error
and causing a deadlock
* Add additional telemetry in order to understand which data sources are the most useful
* Make actions side by side in the configuration page
* Restore main run command
* Update links to point to wiki pages for Github, Notion integrations
* Stanardize nomenclature of the api_type to use _config suffix
Remove header fields that aren't actually helpful for understanding config usage
- Memory consumption now only scales with search models used, not with
content types as well. Previously each content type had it's own
copy of the search ML models. That'd result in 300+ Mb per enabled
content type
- Split model state into 2 separate state objects, `search_models' and
`content_index'.
This allows loading text_search and image_search models first and then
reusing them across all content_types in content_index
- This should cut down memory utilization quite a bit for most users.
I see a ~50% drop in memory utilization.
This will, of course, vary for each user based on the amount of
content indexed vs number of plugins enabled
- This does not solve the RAM utilization scaling with size of the index.
As the whole content index is still kept in RAM while Khoj is running
Should help with #195, #301 and #303
My account doesn't have gpt-4 enabled and it wouldn't work as the default value was always used from extract_questions, where the caller could use the configured model.
- Provide more details on what clicking configure, initialize buttons
or changing the results count slider does
- This shows up on user hovering over those buttons
* For the demo instance, re-instate the scheduler, but infrequently for api updates
- In constants, determine the cadence based on whether it's a demo instance or not
- This allow us to collect telemetry again. This will also allow us to save the chat session
* Conditionally skip updating the index altogether if it's a demo isntance
* Add backend support for Notion data parsing
- Add a NotionToJsonl class which parses the text of Notion documents made accessible to the API token
- Make corresponding updates to the default config, raw config to support the new notion addition
* Add corresponding views to support configuring Notion from the web-based settings page
- Support backend APIs for deleting/configuring notion setup as well
- Streamline some of the index updating code
* Use defaults for search and chat queries results count
* Update pagination of retrieving pages from Notion
* Update state conversation processor when update is hit
* frequency_penalty should be passed to gpt through kwargs
* Add check for notion in render_multiple method
* Add headings to Notion render
* Revert results count slider and split Notion files by blocks
* Clean/fix misc things in the function to update index
- Use the successText and errorText variables appropriately
- Name parameters in function calls
- Add emojis, woohoo
* Clean up and further modularize code for processing data in Notion
* Add langchain static files and pytorch metadata to Khoj native app
* Add pillow static files, metadata & hidden imports to Khoj native app
* Fix path to web interface static files on Khoj native app
* Add tiktoken hidden imports to make chat work from Khoj native app
* Fix Khoj native app to run with GUI mode enabled
This got broken when we moved from using the --no-gui flag to using
--gui in https://github.com/khoj-ai/khoj/pull/263
* Update the /chat endpoint to conditionally support streaming
- If streams are enabled, return the threadgenerator as it does currently
- If stream is disabled, return a JSON response with the response/compiled references separated out
- Correspondingly, update the chat.html UI to use the streamed API, as well as Obsidian
- Rename chat/init/ to chat/history
* Update khoj.el to use the /history endpoint
- Update corresponding unit tests to use stream=true
* Remove & from call to /chat for obsidian
* Abstract functions out into a helpers.py file and clean up some of the error-catching
Deprecate usage of the older gpt3 models in-place of the newer chat
based models
- text-davinci-003 is only 50% cheaper than gpt4 and less reliable for
question extraction
- Using gpt-3.50turbo for summarization should reduce cost of chat
- Keep conversation.chat_session as a list instead of a string
- Update completion_with_backoff func to use ChatML format
- Fix testing gpt converse method after it started streaming responses
- Pass stop in model_kwargs dictionary and api key in openai_api_key
parameter to chat completion methods. This should resolve the arg
warning thrown by OpenAI module
The previous json parsing was failing to handle questions with date
filters
Fix the chat actor tests to run without throwing error with freezegun
complaining about importing transformers.local_llama model
Remove quote escapes from date filter examples provided to
extract_questions actor
- Before
Only the search interface had the results count configuration option
- After
- The results count is set on the settings page instead of the
search page
- Both search and chat can use the configured results count instead
of just search
* For the demo instance, re-instate the scheduler, but infrequently for api updates
- In constants, determine the cadence based on whether it's a demo instance or not
- This allow us to collect telemetry again. This will also allow us to save the chat session
* Conditionally skip updating the index altogether if it's a demo isntance
- What
- Stream chat responses from OpenAI API to Web, Obsidian clients
- Implement using a callback function which manages a queue where new tokens can be placed as they come on. As the thread is read from, tokens are removed.
- When the final token has been processed, add the `compiled_references` to the queue to be rendered by the `chat` client
- When the thread has been closed, save the accumulated conversation log in the user's history using a `partial func`
- Incrementally decode tokens on the front end and add them as they appear from the streamed response
- Why
This significantly reduces perceived latency and OpenAI API request timeouts for Chat
Closes https://github.com/khoj-ai/khoj/issues/257
- I needed to installed node-fetch to accomplish this, as the built-in request object from Obsidian doesn't seem to support streaming and the built-in fetch object is very sensitive to any and all cross origin requests
Removing unused content types will reduce khoj code to manage
- 0f993b3 Drop support for Ledger as a separate content type
Khoj will soon get a generic text indexing content type in Index plain text files #237.
This along with a file filter should suffice for searching through Ledger transactions
- c9db532 Remove unused org-music as an indexable content type from Khoj
Org-music was just a custom content type that worked with org-music.
It was mostly only useful for me.
Khoj will soon get a generic text indexing content type. This along
with a file filter should suffice for searching through Ledger
transactions, if required.
Having a specific content type for niche use-case like ledger isn't
useful. Removing unused content types will reduce khoj code to manage.
Org-music was just a custom content type that worked with org-music.
It was mostly only useful for me.
Cleaning up that code will reduce number of content types for khoj to
manage.
- Add one-click disablement
- Remove fields that probably don't need to be edited (our implementation details)
- Add a green tick if a given field is configured
- In theory, this will be suitable for any Khoj instance that's meant for external-facing purposes (as in, outside of the user's network)
- Prevent re-indexing for Github data if this is a demo instance
- Fix up some issues with the CSS which made settings page small in mobile
- In the frontend views for Khoj, add a button to get on the waitlist and links to the landing page
- Break out of rendering list if at end of org block in org.js
- This would previous hang rendering results in web interface
Should try fix this upstream in org.js as well
- Previously Khoj could only support Python upto 3.10 due to pytorch.
But lots of folks had python 3.11 installed by default on their machines.
This required installing python 3.10 and dealing with virtual envs.
With Torch >= 2.0.1 now able to support python 3.11, at least one
class of installation troubles for Khoj should drop. See
https://github.com/pytorch/pytorch/issues/86566 for reference
- Preliminary testing indicates using the new torch 2.x may reduce
search time by 25% (from 80ms to 60ms on Mac M1)
- Update Docs to not require mentioning python <=3.10 required
- Update Github test workflow to run khoj tests with python 3.11 too
- Use a request session to reduce the overhead of setting up a new connection with the Github URL each request
- Use the streaming feature for the REST api to reduce some of the memory footprint
- Set image_search.query to async to use it with multi-threading
This is same as text_search.query being set to an async method
- Exit search early if no search_model is defined in state.model
- So when searching across content types (with content-type = "all")
org-mode results get rendered differently than markdown, PDF etc. results
- Set div class for each result separately instead of a single uber div
for styling. This allows styling div of each result based on the
content-type of that result
- No need to create placeholder "all" content type on web interface as
server is passing an all content type by itself
- Add cards to configure each of the Github repositories
- Fix a bug in the API which caused all other settings to be wiped when updating one of the content types
- Provide an error message to the user if they have a misconfiguration in their chat settings
- Add support for indexing org files as well as markdown files from the Github repository and update corresponding search view
- Support indexing a list of repositories
- Show success/failure status message much closer to the save button
Previously status message was shown on top of the page, which wasn't
always in view and wasn't easily seen
- Improve the status message to more clearly show next steps on success
If no content-type selected in transient menu option, khoj.el queries
khoj server without content-type parameter (t) set.
This results in search across all enabled asymmetric search text
content types
- Add new filter abstract method to remove filter terms from query
- Use the filter method to remove filter terms, encode this defiltered
query and pass it to the query methods of each search types
TODO: Encoding query is still taking 100-200 ms unlike before. Need to
investigate why
- Update API to return content from all enabled content types when type
is not set to specific type in HTTP request param
- To do this efficiently run the search queries in parallel threads
- Default is 30. So number of paginated requests required to get all
items (commits, files) will reduce by 67%
- No need to increase page size for the get tree Github API request from
`get_markdown_files'
Get tree Github API doesn't support pagination and return 100K items
in response. This should be way more than enough for our current
use-cases
- Previously wasn't prefixing "token" to PAT token in Auth header
This resulted in the request being considered unauthenticated
- Unauthenticated requests to Github API are limited to 60 requests/hour
Authenticated requests to Github API are allowed 5000 requests/hour
- Add a central configuration management page to make management of config details easier
- Add relevant api endpoints both for client and server to update/request data as necessary
- Attempt to update the favicon
The Llama_Hub Github plugin is fairly limited.
The Github Rest API is well supported and can easily be extended to
index commit messages, issues, discussions, PRs etc.
- Make API endpoints on Khoj server accept `client` as request parameter
- Khoj API endpoints: /chat, /search, /update
- Make Khoj clients set `client` request param when calling the API endpoints on the Khoj server
- Khoj clients: Emacs, Obsidian and Web
- Also log khoj server_version running to telemetry server
- This improves latency of @general chat by avoiding unnecessary
compute
- It also avoids passing references in API response when they haven't
been used to generate the chat response. So interfaces don't have to
add logic to not render them unnecessarily
- Make plugin update khoj server config to index PDF files in vault too
- Make Obsidian plugin update index for PDF files in vault too
- Show PDF results in Khoj Search modal as well
- Ensure combined results are sorted by score across both types
- Jump to PDF file when select it PDF search result from modal
- Match argument names passed to khoj openai completion funcs with
arguments passed to langchain calls to OpenAI
- This simplifies the logic in the khoj openai completion funcs
- Fix bug where both LangChain and Khoj retry requests 6 times each.
So a total of 12 requests at >1minute intervals for each chat
response in case of OpenAI API being down
- Retrying too many times when the API is failing doesn't help
- The earlier 60 second request timeout was spacing out the interval
between retries way too much. This slowed down chat response times
quite a bit when API was being flaky
- With these updates you'll know if call to chat API failed in under a
minute
- Use ChatModel and ChatOpenAI to call OpenAI chat model instead of
using OpenAI package directly
- This is being done as part of migration to rely on LangChain for
creating agents and managing their state
