The query method had become too big.
Extract out filter, score, sort and deduplicate logic used by
text_search.query into separate methods.
This should improve readabilty of code.
- Changes
- Fix method signatures of BaseFilter subclasses.
Else typing information isn't translating to them
- Explicitly pass `entries: list[Entry]' as arg to `load' method
- Fix type of `raw_entries' arg to `apply' method
to list[Entry] from list[str]
- Rename `raw_entries' arg to `apply' method to `entries'
- Fix `raw_query' arg used in `apply' method of subclasses to `query'
- Set type of entries, corpus_embeddings in TextSearchModel
- Verification
Ran `mypy --config-file .mypy.ini src' to verify typing
- `torch.Tensor' is apparently a legacy tensor constructor
- Using that to create tensor on MPS devices throws error:
RuntimeError: legacy constructor expects device type: cpu but device type: mps was passed
- `torch.tensor' can handle creating tensors on Mac GPU (MPS) fine
This is unlike the more general chat API that combines summarization
of top search result and conversing with the OpenAI model
This should give faster summary results. As no intent categorization
API call required
- Use latest davinci model for tests
- Wrap prompt in triple quotes to improve legibilty
- `understand' method returns dictionary instead of string. Fix its test
- Fix prompt for new model to pass `chat_with_history' test
- Default to using `text-davinci-003' if conversation model not
explicitly configured by user. Stop using the older `davinci' and
`davinci-instruct' models
- Use `model' instead of `engine' as parameter.
Usage of `engine' parameter in OpenAI API is deprecated
- Init processor before search to instantiate `openai_api_key'
from `khoj.yml'. The key is used to configure search with openai models
- To use OpenAI models for search in Khoj
- Set `encoder' to name of an OpenAI model. E.g text-embedding-ada-002
- Set `encoder-type' in `khoj.yml' to `src.utils.models.OpenAI'
- Set `model-directory' to `null', as online model cannot be stored on disk
Long words (>500 characters) provide less useful context to models.
Dropping very long words allow models to create better embeddings by
passing more of the useful context from the entry to the model
- Previously `model_type' was set in the setup of each `search_type'
- All encoders were of type `SentenceTransformer'
- All cross_encoders were of type `CrossEncoder'
- Now `encoder-type' can be configured via the new `encoder_type' field
in `TextSearchConfig' under `search-type` in `khoj.yml`.
- All the specified `encoder-type' class needs is an `encode' method
that takes entries and returns embedding vectors
- Ensure all tensors are on MPS device before doing operations across them
- Background
- GPU is used by default for Khoj on MacOS now
- Needed PyTorch > 1.13.0 on Macs to use GPU, which we do now
- MPS should speed up search and indexing on MacOS
Fix usage warning for unescaped single quote in `khoj.el' docstring.
Converts usage of '<text>' into `<text>' to use the correct quote forms in generated docs
⛔ Warning (comp): khoj.el:119:2: Warning: docstring has wrong usage of unescaped single quotes (use \= or different quoting)
⛔ Warning (comp): khoj.el:120:2: Warning: docstring has wrong usage of unescaped single quotes (use \= or different quoting)
⛔ Warning (comp): khoj.el:121:2: Warning: docstring has wrong usage of unescaped single quotes (use \= or different quoting)
⛔ Warning (comp): khoj.el:168:2: Warning: docstring has wrong usage of unescaped single quotes (use \= or different quoting)
- Features
- Search using Khoj from within the Obsidian app
Allow Natural language search on your (markdown) notes in Obsidian Vault
- Show search results as rendered (instead of raw) Markdown
Improve legibility of the results
- Jump to selected note from search result in Khoj search modal
Simplify seeing result within its original note context
- Automatically configure khoj to index markdown files in current vault
Reduce khoj setup steps for plugin users by using reasonable defaults
- Code updates the markdown config in khoj.yml and triggers index update
- It can be configured by user in khoj plugin settings, if required
- Add Demo and detailed Readme for the Obsidian plugin
Ease setup and usage. Give context about capabilities
- Miscellaneous
- Trying keep a mono repo until the Khoj project is mature enough
to reduce maintainance burden
This can ease configuring khoj from the different interfaces
- Don't need to know all the (default) config used by khoj.
- Just get default config by calling the above API endpoint.
- Then modify desired portions and call POST /api/config/data to
configure khoj.
- Start khoj server (in non-GUI mode) without needing config file
already instantiated.
- But throw warning to configure khoj to use it
- This allows plugins to configure the app via the /config/data APIs
- To be used by the Khoj obsidian plugin to configure markdown content
in khoj
- Poll scheduler every minute using threading.Timer
- Use 60 seconds polling interval to avoid fork bombing
- Schedule next via the same poll scheduler
- Allow clean program interrupt by running scheduler in daemon mode
- There are 3 paths to updating/setting the index (stored in state.model)
- App start
- API
- Scheduler
- Put all updates to the index behind a lock. As multiple updates path
that could (potentially) run at the same time (via API or Scheduler)
- Remove property drawer from test entry for max_words splitting test
- Property drawer is not required for the test
- Keep minimal test case to reduce chance for confusion
- Required because entries are now split by the max_word count supported
by the ML models
- This would now result in potentially duplicate hits, entries being
returned to user
- Do deduplication after ranking to get the top ranked deduplicated
results
- The instructions suggest installing khoj-assistant via pip install.
This installs the latest tagged/release version of khoj
- To match that version user should install khoj.el from MELPA stable
instead of MELPA
- Issue
ML Models truncate entries exceeding some max token limit.
This lowers the quality of search results
- Fix
Split entries by max tokens before indexing.
This should improve searching for content in longer entries.
- Miscellaneous
- Test method to split entries by max tokens
Update readme to ask user to install khoj.el from MELPA when a
pre-release version of the main khoj app is installed. Else install
khoj.el from MELPA Stable
- Reason
- All clients that currently consume the API are part of Khoj
- Any breaking API changes will be fixed in clients immediately
- So decoupling client from API is not required
- This removes the burden of maintaining muliple versions of the API
- Context
- The app maintains all text content in a standard, intermediate format
- The intermediate format was loaded, passed around as a dictionary
for easier, faster updates to the intermediate format schema initially
- The intermediate format is reasonably stable now, given it's usage
by all 3 text content types currently implemented
- Changes
- Concretize text entries into `Entries' class instead of using dictionaries
- Code is updated to load, pass around entries as `Entries' objects
instead of as dictionaries
- `text_search' and `text_to_jsonl' methods are annotated with
type hints for the new `Entries' type
- Code and Tests referencing entries are updated to use class style
access patterns instead of the previous dictionary access patterns
- Move `mark_entries_for_update' method into `TextToJsonl' base class
- This is a more natural location for the method as it is only
(to be) used by `text_to_jsonl' classes
- Avoid circular reference issues on importing `Entries' class
- Both Text, Image Search were already giving list of entry, score
- This change just concretizes this change and exposes this in the API
documentation (i.e OpenAPI, Swagger, Redocs)
- Split router.py into v1.0, beta and frontend (no-prefix) api modules
under new router package. Version tag in main.py via prefix
- Update frontends to use the versioned api endpoints
- Update tests to work with versioned api endpoints
- Update docs to mentioned, reference only versioned api endpoints
In my installation, it appears that `url-request-method` is sometimes set
globally to POST. Need to explicitly set it to ensure that GET is always
used as intended.
- Pass force=true to /update API to force regenerating index from
scratch
- Otherwise calls to the /update API endpoint will result in an
incremental update to index
- Start standardizing implementation of the `text_to_jsonl' processors
- `text_to_jsonl; scripts already had a shared structure
- This change starts to codify that implicit structure
- Benefits
- Ease adding more `text_to_jsonl; processors
- Allow merging shared functionality
- Help with type hinting
- Drawbacks
- Lower agility to change. But this was already an implicit issue as
the text_to_jsonl processors got more deeply wired into the app
- Pillow already supports reading XMP metadata from Images
- Removes need to maintain my fork of unmaintained PyExiftool
- This also removes dependency on system Exiftool package for
XMP metadata extraction
- Add test to verify XMP metadata extracted from test images
- Remove references to Exiftool from Documentation
- Simplify tracking khoj query history, saving/sharing links
- Do not execute search, when query only contains whitespaces
- Prevents error when try process results of empty query
- As `/reload` updates index incrementally, it's relatively quick
- This makes exposing `/reload` endpoint a better default to expose
via the web interface than `the /regenerate' endpoint
- For queries with only filters in them short-circuit and return
filtered results. No need to run semantic search, re-ranking.
- Add client test for filter only query and quote query in client tests
- Image search already uses a sorted list of images to process
- Prevents index of entries to desync when entries, embeddings
generated by a separate server/app instance
- Update existings code, tests to process input-filters as list
instead of str
- Test `text_to_jsonl' get files methods to work with combination of
`input-files' and `input-filters'
Resolves#84
- Provides more control to invalidate cache on update to entries, embeddings
- Allows logging when results are being returned from cache etc
- FastAPI, Swagger API docs look better as the `search' controller not
wrapped in generically named function when using functools LRU decorator
- Issue
- Indent regex was previously catching escape sequences like newlines
- This was resulting in entries with only escape sequences in body to
be prepended to property drawers etc during rendering
- Fix
- Update indent regex to only look for spaces in each line
- Only render body when body contains non-escape characters
- Create test to prevent this regression from silently resurfacing
- Previously heading entries were not indexed to maintain search quality
- But given that there are use-cases for indexing entries with no body
- Add a configurable `index_heading_entries' field to index heading entries
- This `TextContentConfig' field is currently only used for OrgMode content
- Let the specific text_to_jsonl method decide which of the
TextContentConfig fields it needs to convert <text> type to jsonl
- This simplifies extending TextContentConfig for a specific type without
modifying all text_to_jsonl methods
- It keeps the number of args being passed to the `text_to_jsonl'
methods in check
- It's more of a hassle to not let word filter go stale on entry
updates
- Generating index on 120K lines of notes takes 1s. Loading from file
takes 0.2s. For less content load time difference will be even smaller
- Let go of startup time improvement for simplicity for now
- Comparing compiled entries is the appropriately narrow target to
identify entries that need to encode their embedding vectors. Given we
pass the compiled form of the entry to the model for encoding
- Hashing the whole entry along with it's raw form was resulting in a
bunch of entries being marked for updated as LINE: <entry_line_no>
is a string added to each entries raw format.
- This results in an update to a single entry resulting in all entries
below it in the file being marked for update (as all their line
numbers have changed)
- Log performance metrics for steps to convert org entries to jsonl
- Having Tags as sets was returning them in a different order
everytime
- This resulted in spuriously identifying existing entries as new
because their tags ordering changed
- Converting tags to list fixes the issue and identifies updated new
entries for incremental update correctly
- What
- Hash the entries and compare to find new/updated entries
- Reuse embeddings encoded for existing entries
- Only encode embeddings for updated or new entries
- Merge the existing and new entries and embeddings to get the updated
entries, embeddings
- Why
- Given most note text entries are expected to be unchanged
across time. Reusing their earlier encoded embeddings should
significantly speed up embeddings updates
- Previously we were regenerating embeddings for all entries,
even if they had existed in previous runs
