- Benefits of moving to llama-cpp-python from gpt4all:
- Support for all GGUF format chat models
- Support for AMD, Nvidia, Mac, Vulcan GPU machines (instead of just Vulcan, Mac)
- Supports models with more capabilities like tools, schema
enforcement, speculative ddecoding, image gen etc.
- Upgrade default chat model, prompt size, tokenizer for new supported
chat models
- Load offline chat model when present on disk without requiring internet
- Load model onto GPU if not disabled and device has GPU
- Load model onto CPU if loading model onto GPU fails
- Create helper function to check and load model from disk, when model
glob is present on disk.
`Llama.from_pretrained' needs internet to get repo info from
HuggingFace. This isn't required, if the model is already downloaded
Didn't find any existing HF or llama.cpp method that looked for model
glob on disk without internet
* Initial pass at backend changes to support agents
- Add a db model for Agents, attaching them to conversations
- When an agent is added to a conversation, override the system prompt to tweak the instructions
- Agents can be configured with prompt modification, model specification, a profile picture, and other things
- Admin-configured models will not be editable by individual users
- Add unit tests to verify agent behavior. Unit tests demonstrate imperfect adherence to prompt specifications
* Customize default behaviors for conversations without agents or with default agents
* Add a new web client route for viewing all agents
* Use agent_id for getting correct agent
* Add web UI views for agents
- Add a page to view all agents
- Add slugs to manage agents
- Add a view to view single agent
- Display active agent when in chat window
- Fix post-login redirect issue
* Fix agent view
* Spruce up the 404 page and improve the overall layout for agents pages
* Create chat actor for directly reading webpages based on user message
- Add prompt for the read webpages chat actor to extract, infer
webpage links
- Make chat actor infer or extract webpage to read directly from user
message
- Rename previous read_webpage function to more narrow
read_webpage_at_url function
* Rename agents_page -> agent_page
* Fix unit test for adding the filename to the compiled markdown entry
* Fix layout of agent, agents pages
* Merge migrations
* Let the name, slug of the default agent be Khoj, khoj
* Fix chat-related unit tests
* Add webpage chat command for read web pages requested by user
Update auto chat command inference prompt to show example of when to
use webpage chat command (i.e when url is directly provided in link)
* Support webpage command in chat API
- Fallback to use webpage when SERPER not setup and online command was
attempted
- Do not stop responding if can't retrieve online results. Try to
respond without the online context
* Test select webpage as data source and extract web urls chat actors
* Tweak prompts to extract information from webpages, online results
- Show more of the truncated messages for debugging context
- Update Khoj personality prompt to encourage it to remember it's capabilities
* Rename extract_content online results field to webpages
* Parallelize simple webpage read and extractor
Similar to what is being done with search_online with olostep
* Pass multiple webpages with their urls in online results context
Previously even if MAX_WEBPAGES_TO_READ was > 1, only 1 extracted
content would ever be passed.
URL of the extracted webpage content wasn't passed to clients in
online results context. This limited them from being rendered
* Render webpage read in chat response references on Web, Desktop apps
* Time chat actor responses & chat api request start for perf analysis
* Increase the keep alive timeout in the main application for testing
* Do not pipe access/error logs to separate files. Flow to stdout/stderr
* [Temp] Reduce to 1 gunicorn worker
* Change prod docker image to use jammy, rather than nvidia base image
* Use Khoj icon when Khoj web is installed on iOS as a PWA
* Make slug required for agents
* Simplify calling logic and prevent agent access for unauthenticated users
* Standardize to use personality over tuning in agent nomenclature
* Make filtering logic more stringent for accessible agents and remove unused method:
* Format chat message query
---------
Co-authored-by: Debanjum Singh Solanky <debanjum@gmail.com>
### Overview
Khoj can now read website directly without needing to go through the search step first
### Details
- Parallelize simple webpage read and extractor
- Rename extract_content online results field to web pages
- Tweak prompts to extract information from webpages, online results
- Test select webpage as data source and extract web urls chat actors
- Render webpage read in chat response references on Web, Desktop apps
- Pass multiple webpages with their urls in online results context
- Support webpage command in chat API
- Add webpage chat command for read web pages requested by user
- Create chat actor for directly reading webpages based on user message
* Initial pass at backend changes to support agents
- Add a db model for Agents, attaching them to conversations
- When an agent is added to a conversation, override the system prompt to tweak the instructions
- Agents can be configured with prompt modification, model specification, a profile picture, and other things
- Admin-configured models will not be editable by individual users
- Add unit tests to verify agent behavior. Unit tests demonstrate imperfect adherence to prompt specifications
* Customize default behaviors for conversations without agents or with default agents
* Use agent_id for getting correct agent
* Merge migrations
* Simplify some variable definitions, add additional security checks for agents
* Rename agent.tuning -> agent.personality
Previously was assuming the system prompt is being always passed as
the first message. So expected there to be at least 2 messages in logs.
This broke chat actors querying with single long non system message.
A more robust way to extract system prompt is via the message role
instead
- Add prompt for the read webpages chat actor to extract, infer
webpage links
- Make chat actor infer or extract webpage to read directly from user
message
- Rename previous read_webpage function to more narrow
read_webpage_at_url function
- Remove stale tests
- Improve tests to pass across gpt-3.5 and gpt-4-turbo
- The haiku creation director was failing because of duplicate query in
instantiated prompt
- Time reading webpage, extract info from webpage steps for perf
analysis
- Deduplicate webpages to read gathered across separate google
searches
- Use aiohttp to make API requests non-blocking, pair with asyncio to
parallelize all the online search webpage read and extract calls
* Make major improvements to the image generation flow
- Include user context from online references and personal notes for generating images
- Dynamically select the modality that the LLM should respond with
- Retun the inferred context in the query response for the dekstop, web chat views to read
* Add unit tests for retrieving response modes via LLM
* Move output mode unit tests to the actor suite, rather than director
* Only show the references button if there is at least one available
* Rename aget_relevant_modes to aget_relevant_output_modes
* Use a shared method for generating reference sections, simplify some of the prompting logic
* Make out of space errors in the desktop client more obvious
* Display given_name field only if it is not None
* Add default slugs in the migration script
* Ensure that updated_at is saved appropriately, make sure most recent chat is returned for default history
* Remove the bin button from the chat interface, given deletion is handled in the drop-down menus
* Refresh the side panel when a new chat is created
* Improveme tool retrieval prompt, don't let /online fail, and improve parsing of extract questions
* Fix ending chat response by offline chat on hitting a stop phrase
Previously the whole phrase wouldn't be in the same response chunk, so
chat response wouldn't stop on hitting a stop phrase
Now use a queue to keep track of last 3 chunks, and to stop responding
when hit a stop phrase
* Make chat on Obsidian backward compatible post chat session API updates
- Make chat on Obsidian get chat history from
`responseJson.response.chat' when available (i.e when using new api)
- Else fallback to loading chat history from
responseJson.response (i.e when using old api)
* Fix detecting success of indexing update in khoj.el
When khoj.el attempts to index on a Khoj server served behind an https
endpoint, the success reponse status contains plist with certs. This
doesn't mean the update failed.
Look for :errors key in status instead to determine if indexing API
call failed. This fixes detecting indexing API call success on the
Khoj Emacs client, even for Khoj servers running behind SSL/HTTPS
* Fix the mechanism for populating notes references in the conversation primer for both offline and online chat
* Return conversation.default when empty list for dynamic prompt selection, send all cmds in telemetry
* Fix making chat on Obsidian backward compatible post chat session API updates
New API always has conversation_id set, not `chat' which can be unset
when chat session is empty.
So use conversation_id to decide whether to get chat logs from
`responseJson.response.chat' or `responseJson.response' instead
---------
Co-authored-by: Debanjum Singh Solanky <debanjum@gmail.com>
* Have Khoj dynamically select which conversation command(s) are to be used in the chat flow
- Intercept the commands if in default mode, and have Khoj dynamically guess which tools would be the most relevant for answering the user's query
* Remove conditional for default to enter online search mode
* Add multiple-tool examples in the prompt, make prompt for tools more specific to info collection
- Can now expect date awareness chat quality test to pass
- Prevent offline chat model from printing verbatim user Notes and
special tokens
- Make it ask follow-up questions if it needs more context
* Initailize changes to incporate web scraping logic after getting SERP results
- Do some minor refactors to pass a symptom prompt to the openai model when making a query
- integrate Olostep in order to perform the webscraping
* Fix truncation error with new line, fix typing in olostep code
* Use the authorization header for the token
* Add a small hint/indicator for how to use Khojs other modalities in the welcome prompt
* Add more detailed error message if Olostep query fails
* Add unit tests which invoke Olostep in chat director
* Add test for olostep tool
### Major
- Short-circuit API rate limiter for unauthenticated user
Calls by unauthenticated users were failing at API rate limiter as it
failed to access user info object. This is a bug.
API rate limiter should short-circuit for unauthenicated users so a
proper Forbidden response can be returned by API
Add regression test to verify that unauthenticated users get 403
response when calling the /chat API endpoint
### Minor
- Remove trailing slash to normalize khoj url in obsidian plugin settings
- Move used /api/config API controllers into separate module
- Delete unused /api/beta API endpoint
- Fix error message rendering in khoj.el, khoj obsidian chat
- Handle deprecation warnings for subscribe renew date, langchain, pydantic & logger.warn
- Ensure langchain less than 0.2.0 is used, to prevent breaking
ChatOpenAI, PyMuPDF usage due to their deprecation after 0.2.0
- Set subscription renewal date to a timezone aware datetime
- Use logger.warning instead of logger.warn as latter is deprecated
- Use `model_dump' not deprecated dict to get all configured content_types
Calls by unauthenticated users were failing at API rate limiter as it
failed to access user info object. This is a bug.
API rate limiter should short-circuit for unauthenicated users so a
proper Forbidden response can be returned by API
Add regression test to verify that unauthenticated users get 403
response when calling the /chat API endpoint
- All search models are loaded into memory, and stored in a dictionary indexed by name
- Still need to add database migrations and create a UI for user to select their choice. Presently, it uses the default option
- Add a dependency on the indexer API endpoint that rounds up the amount of data indexed and uses that to determine whether the next set of data should be processed
- Delete any files that are being removed for adminstering the calculation
- Show current amount of data indexed in the config page
- Our pypi package currently does not work because the django app and associated database is not included. To remedy this issue, move the app into the src/khoj folder. This has the added benefit of improved organization of the codebase, as all server related code is now in a single folder
- Update associated file paths and system references
### Overview
The parent hierarchy of org-mode entries can store important context.
This change updates OrgNode to track parent headings for each org entry and adds the parent outline for each entry to the index
### Details
- Test search uses ancestor headings as context for improved results
- Add ancestor headings of each org-mode entry to their compiled form
- Track ancestor headings for each org-mode entry in org-node parser
Resolves#85
- Upgrade FastAPI to >= latest version. Required upgrade of FastAPI.
Earlier version didn't support wrapping common query params in class
- Use per fixture app instead of a global FastAPI app in conftest
- Upgrade minimum required Django version
- Fix no notes chat director test with updated no notes message
No notes message was updated in commit 118f1143
- Update test data to add deeper outline hierarchy for testing
hierarchy as context
- Update collateral tests that need count of entries updated, deleted
asserts to be updated
- Expose ability to modify search model via Django admin interface
- Previously the bi_encoder and cross_encoder models to use were set
in code
- Now it's user configurable but with a default config generated by
default
- Notes prompt doesn't need to be so tuned to question answering. User
could just want to talk about life. The notes need to be used to
response to those, not necessarily only retrieve answers from notes
- System and notes prompts were forcing asking follow-up questions a
little too much. Reduce strength of follow-up question asking
This will be useful for updating, deleting entries by their data
source. Data source can be one of Computer, Github or Notion for now
Store each file/entries source in database
Major
- Ensure search results logic consistent across migration to DB, multi-user
- Manually verified search results for sample queries look the same across migration
- Flatten indexing code for better indexing progress tracking and code readability
Minor
- a4f407f Test memory leak on MPS device when generating vector embeddings
- ef24485 Improve Khoj with DB setup instructions in the Django app readme (for now)
- f212cc7 Arrange remaining text search tests in arrange, act, assert order
- 022017d Fix text search tests to test updated indexing log messages
- Rather than having each individual user configure their conversation settings, allow the server admin to configure the OpenAI API key or offline model once, and let all the users re-use that code.
- To configure the settings, the admin should go to the `django/admin` page and configure the relevant chat settings. To create an admin, run `python3 src/manage.py createsuperuser` and enter in the details. For simplicity, the email and username should match.
- Remove deprecated/unnecessary endpoints and views for configuring per-user chat settings
Improves readability as name has closer match to underlying
constructs
- Entry is any atomic item indexed by Khoj. This can be an org-mode
entry, a markdown section, a PDF or Notion page etc.
- Embeddings are semantic vectors generated by the search ML model
that encodes for meaning contained in an entries text.
- An "Entry" contains "Embeddings" vectors but also other metadata
about the entry like filename etc.
- Add a productionized setup for the Khoj server using `gunicorn` with multiple workers for handling requests
- Add a new Dockerfile meant for production config at `ghcr.io/khoj-ai/khoj:prod`; the existing Docker config should remain the same
### ✨ New
- Use API keys to authenticate from Desktop, Obsidian, Emacs clients
- Create API, UI on web app config page to CRUD API Keys
- Create user API keys table and functions to CRUD them in Database
### 🧪 Improve
- Default to better search model, [gte-small](https://huggingface.co/thenlper/gte-small), to improve search quality
- Only load chat model to GPU if enough space, throw error on load failure
- Show encoding progress, truncate headings to max chars supported
- Add instruction to create db in Django DB setup Readme
### ⚙️ Fix
- Fix error handling when configure offline chat via Web UI
- Do not warn in anon mode about Google OAuth env vars not being set
- Fix path to load static files when server started from project root
- Add a data model which allows us to store Conversations with users. This does a minimal lift over the current setup, where the underlying data is stored in a JSON file. This maintains parity with that configuration.
- There does _seem_ to be some regression in chat quality, which is most likely attributable to search results.
This will help us with #275. It should become much easier to maintain multiple Conversations in a given table in the backend now. We will have to do some thinking on the UI.
- Make most routes conditional on authentication *if anonymous mode is not enabled*. If anonymous mode is enabled, it scaffolds a default user and uses that for all application interactions.
- Add a basic login page and add routes for redirecting the user if logged in
- Partition configuration for indexing local data based on user accounts
- Store indexed data in an underlying postgres db using the `pgvector` extension
- Add migrations for all relevant user data and embeddings generation. Very little performance optimization has been done for the lookup time
- Apply filters using SQL queries
- Start removing many server-level configuration settings
- Configure GitHub test actions to run during any PR. Update the test action to run in a containerized environment with a DB.
- Update the Docker image and docker-compose.yml to work with the new application design
GPT4all now supports gguf llama.cpp chat models. Latest
GPT4All (+mistral) performs much at least 3x faster.
On Macbook Pro at ~10s response start time vs 30s-120s earlier.
Mistral is also a better chat model, although it hallucinates more
than llama-2
Ignore .org, .pdf etc. suffixed directories under `input-filter' from
being evaluated as files.
Explicitly filter results by input-filter globs to only index files,
not directory for each text type
Add test to prevent regression
Closes#448
On Windows, the default locale isn't utf8. Khoj had regressed to
reading files in OS specified locale encoding, e.g cp1252, cp949 etc.
It now explicitly uses utf8 encoding to read text files for indexing
Resolves#495, resolves#472
### Overview
- Add ability to push data to index from the Emacs, Obsidian client
- Switch to standard mechanism of syncing files via HTTP multi-part/form. Previously we were streaming the data as JSON
- Benefits of new mechanism
- No manual parsing of files to send or receive on clients or server is required as most have in-built mechanisms to send multi-part/form requests
- The whole response is not required to be kept in memory to parse content as JSON. As individual files arrive they're automatically pushed to disk to conserve memory if required
- Binary files don't need to be encoded on client and decoded on server
### Code Details
### Major
- Use multi-part form to receive files to index on server
- Use multi-part form to send files to index on desktop client
- Send files to index on server from the khoj.el emacs client
- Send content for indexing on server at a regular interval from khoj.el
- Send files to index on server from the khoj obsidian client
- Update tests to test multi-part/form method of pushing files to index
#### Minor
- Put indexer API endpoint under /api path segment
- Explicitly make GET request to /config/data from khoj.el:khoj-server-configure method
- Improve emoji, message on content index updated via logger
- Don't call khoj server on khoj.el load, only once khoj invoked explicitly by user
- Improve indexing of binary files
- Let fs_syncer pass PDF files directly as binary before indexing
- Use encoding of each file set in indexer request to read file
- Add CORS policy to khoj server. Allow requests from khoj apps, obsidian & localhost
- Update indexer API endpoint URL to` index/update` from `indexer/batch`
Resolves#471#243
New URL query params, `force' and `t' match name of query parameter in
existing Khoj API endpoints
Update Desktop, Obsidian and Emacs client to call using these new API
query params. Set `client' query param from each client for telemetry
visibility
New URL follows action oriented endpoint naming convention used for
other Khoj API endpoints
Update desktop, obsidian and emacs client to call this new API
endpoint
This provides flexibility to use non 1st party supported chat models
- Create migration script to update khoj.yml config
- Put `enable_offline_chat' under new `offline-chat' section
Referring code needs to be updated to accomodate this change
- Move `offline_chat_model' to `chat-model' under new `offline-chat' section
- Put chat `tokenizer` under new `offline-chat' section
- Put `max_prompt' under existing `conversation' section
As `max_prompt' size effects both openai and offline chat models
- Format extract questions prompt format with newlines and whitespaces
- Make llama v2 extract questions prompt consistent
- Remove empty questions extracted by offline extract_questions actor
- Update implicit qs extraction unit test for offline search actor
Instead of using the previous method to push data as json payload of POST request
pass it as files to upload via the multi-part/form to the batch indexer API endpoint
- GPT4All integration had ceased working with 0.1.7 specification. Update to use 1.0.12. At a later date, we should also use first party support for llama v2 via gpt4all
- Update the system prompt for the extract_questions flow to add start and end date to the yesterday date filter example.
- Update all setup data in conftest.py to use new client-server indexing pattern
* Remove GPT4All dependency in pyproject.toml and use multiplatform builds in the dockerization setup in GH actions
* Move configure_search method into indexer
* Add conditional installation for gpt4all
* Add hint to go to localhost:42110 in the docs. Addresses #477
* Remove PySide, gui option from code
* Remove pyside 6 dependency from code
* Remove workflows which build desktop applications
* Update unit tests and update line in documentation
* Remove additional references to pyinstaller, gui
* Add uninstall steps to normal uninstall instructions
* Initial version - setup a file-push architecture for generating embeddings with Khoj
* Use state.host and state.port for configuring the URL for the indexer
* Fix parsing of PDF files
* Read markdown files from streamed data and update unit tests
* On application startup, load in embeddings from configurations files, rather than regenerating the corpus based on file system
* Init: refactor indexer/batch endpoint to support a generic file ingestion format
* Add features to better support indexing from files sent by the desktop client
* Initial commit with Electron application
- Adds electron app
* Add import for pymupdf, remove import for pypdf
* Allow user to configure khoj host URL
* Remove search type configuration from index.html
* Use v1 path for current indexer routes
* Initial version - setup a file-push architecture for generating embeddings with Khoj
* Update unit tests to fix with new application design
* Allow configure server to be called without regenerating the index; this no longer works because the API for indexing files is not up in time for the server to send a request
* Use state.host and state.port for configuring the URL for the indexer
* On application startup, load in embeddings from configurations files, rather than regenerating the corpus based on file system
* Store conversation command options in an Enum
* Move to slash commands instead of using @ to specify general commands
* Calculate conversation command once & pass it as arg to child funcs
* Add /notes command to respond using only knowledge base as context
This prevents the chat model to try respond using it's general world
knowledge only without any references pulled from the indexed
knowledge base
* Test general and notes slash commands in openai chat director tests
---------
Co-authored-by: Debanjum Singh Solanky <debanjum@gmail.com>
* Store conversation command options in an Enum
* Move to slash commands instead of using @ to specify general commands
* Calculate conversation command once & pass it as arg to child funcs
* Add /notes command to respond using only knowledge base as context
This prevents the chat model to try respond using it's general world
knowledge only without any references pulled from the indexed
knowledge base
* Test general and notes slash commands in openai chat director tests
* Update gpt4all tests to use md configuration
* Add a /help tooltip
* Add dynamic support for describing slash commands. Remove default and treat notes as the default type
---------
Co-authored-by: sabaimran <narmiabas@gmail.com>
* Allow indexing to continue even if there's an issue parsing a particular org file
* Use approximation in pytorch comparison in text_search UT, skip additional file parser errors for org files
* Change error of expected failure
* Add support for indexing plaintext files
- Adds backend support for parsing plaintext files generically (.html, .txt, .xml, .csv, .md)
- Add equivalent frontend views for setting up plaintext file indexing
- Update config, rawconfig, default config, search API, setup endpoints
* Add a nifty plaintext file icon to configure plaintext files in the Web UI
* Use generic glob path for plaintext files. Skip indexing files that aren't in whitelist
* 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
Asymmetric search is the only search type used now in khoj.el. So
making distinction between between symmetric and asymmetric search
isn't necessary anymore
* 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
OpenAI conversation processor schema had updated but conftest hadn't
been updated to reflect the same.
Update conftest setup of conversation processor to fix this
* 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
Ensure order of new embedding insertion on incremental update
does not affect the order and value of existing embeddings when
normalization is turned off
Asymmetric was older name used to differentiate between symmetric,
asymmetric search.
Now that text search just uses asymmetric search stick to simpler name
- Current incorrect behavior:
All entries with duplicate compiled form are kept on regenerate
but on update only the last of the duplicated entries is kept
This divergent behavior is not ideal to prevent index corruption
across reconfigure and update
* 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
- 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
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.
