## New EmbeddingStore (metadata) `Filter` API
Many embedding stores, such as
[Pinecone](https://docs.pinecone.io/docs/metadata-filtering) and
[Milvus](https://milvus.io/docs/boolean.md) support strict filtering
(think of an SQL "WHERE" clause) during similarity search.
So, if one has an embedding store with movies, for example, one could
search not only for the most semantically similar movies to the given
user query but also apply strict filtering by metadata fields like year,
genre, rating, etc. In this case, the similarity search will be
performed only on those movies that match the filter expression.
Since LangChain4j supports (and abstracts away) many embedding stores,
there needs to be an embedding-store-agnostic way for users to define
the filter expression.
This PR introduces a `Filter` interface, which can represent both simple
(e.g., `type = "documentation"`) and composite (e.g., `type in
("documentation", "tutorial") AND year > 2020`) filter expressions in an
embedding-store-agnostic manner.
`Filter` currently supports the following operations:
- Comparison:
- `IsEqualTo`
- `IsNotEqualTo`
- `IsGreaterThan`
- `IsGreaterThanOrEqualTo`
- `IsLessThan`
- `IsLessThanOrEqualTo`
- `IsIn`
- `IsNotIn`
- Logical:
- `And`
- `Not`
- `Or`
These operations are supported by most embedding stores and serve as a
good starting point. However, the list of operations will expand over
time to include other operations (e.g., `Contains`) supported by
embedding stores.
Currently, the DSL looks like this:
```java
Filter onlyDocs = metadataKey("type").isEqualTo("documentation");
Filter docsAndTutorialsAfter2020 = metadataKey("type").isIn("documentation", "tutorial").and(metadataKey("year").isGreaterThan(2020));
// or
Filter docsAndTutorialsAfter2020 = and(
metadataKey("type").isIn("documentation", "tutorial"),
metadataKey("year").isGreaterThan(2020)
);
```
## Filter expression as a `String`
Filter expression can also be specified as a `String`. This might be
necessary, for example, if the filter expression is generated
dynamically by the application or by the LLM (as in [self
querying](https://python.langchain.com/docs/modules/data_connection/retrievers/self_query/)).
This PR introduces a `FilterParser` interface with a simple `Filter
parse(String)` API, allowing for future support of multiple syntaxes (if
this will be required).
For the out-of-the-box filter syntax, ANSI SQL's `WHERE` clause is
proposed as a suitable candidate for several reasons:
- SQL is well-known among Java developers
- There is extensive tooling available for SQL (e.g., parsers)
- LLMs are pretty good at generating valid SQL, as there are tons of SQL
queries on the internet, which are included in the LLM training
datasets. There are also specialized LLMs that are trained for
text-to-SQL task, such as [SQLCoder](https://huggingface.co/defog).
The downside is that SQL's `WHERE` clause might not support all
operations and data types that could be supported in the future by
various embedding stores. In such case, we could extend it to a superset
of ANSI SQL `WHERE` syntax and/or provide an option to express filters
in the native syntax of the store.
An out-of-the-box implementation of the SQL `FilterParser` is provided
as a `SqlFilterParser` in a separate module
`langchain4j-embedding-store-filter-parser-sql`, using
[JSqlParser](https://github.com/JSQLParser/JSqlParser) under the hood.
`SqlFilterParser` can parse SQL "SELECT" (or just "WHERE" clause)
statement into a `Filter` object:
- `SELECT * FROM fake_table WHERE userId = '123-456'` ->
`metadataKey("userId").isEqualTo("123-456")`
- `userId = '123-456'` -> `metadataKey("userId").isEqualTo("123-456")`
It can also resolve `CURDATE()` and
`CURRENT_DATE`/`CURRENT_TIME`/`CURRENT_TIMESTAMP`:
`SELECT * FROM fake_table WHERE year = EXTRACT(YEAR FROM CURRENT_DATE`
-> `metadataKey("year").isEqualTo(LocalDate.now().getYear())`
## Changes in `Metadata` API
Until now, `Metadata` supported only `String` values. This PR expands
the list of supported value types to `Integer`, `Long`, `Float` and
`Double`. In the future, more types may be added (if needed).
The method `String get(String key)` will be deprecated later in favor
of:
- `String getString(String key)`
- `Integer getInteger(String key)`
- `Long getLong(String key)`
- etc
New overloaded `put(key, value)` methods are introduced to support more
value types:
- `put(String key, int value)`
- `put(String key, long value)`
- etc
## Changes in `EmbeddingStore` API
New method `search` is added that will become the main entry point for
search in the future. All `findRelevant` methods will be deprecated
later.
New `search` method accepts `EmbeddingSearchRequest` and returns
`EmbeddingSearchResult`.
`EmbeddingSearchRequest` contains all search criteria (e.g.
`maxResults`, `minScore`), including new `Filter`.
`EmbeddingSearchResult` contains a list of `EmbeddingMatch`.
```java
EmbeddingSearchResult search(EmbeddingSearchRequest request);
```
## Changes in `EmbeddingStoreContentRetriever` API
`EmbeddingStoreContentRetriever` can now be configured with a static
`filter` as well as dynamic `dynamicMaxResults`, `dynamicMinScore` and
`dynamicFilter` in the builder:
```java
ContentRetriever contentRetriever = EmbeddingStoreContentRetriever.builder()
.embeddingStore(embeddingStore)
.embeddingModel(embeddingModel)
...
.maxResults(3)
// or
.dynamicMaxResults(query -> 3) // You can define maxResults dynamically. The value could, for example, depend on the query or the user associated with the query.
...
.minScore(0.3)
// or
.dynamicMinScore(query -> 0.3)
...
.filter(metadataKey("userId").isEqualTo("123-456")) // Assuming your TextSegments contain Metadata with key "userId"
// or
.dynamicFilter(query -> metadataKey("userId").isEqualTo(query.metadata().chatMemoryId().toString()))
...
.build();
```
So now you can define `maxResults`, `minScore` and `filter` both
statically and dynamically (they can depend on the query, user, etc.).
These values will be propagated to the underlying `EmbeddingStore`.
##
["Self-querying"](https://python.langchain.com/docs/modules/data_connection/retrievers/self_query/)
This PR also introduces `LanguageModelSqlFilterBuilder` in
`langchain4j-embedding-store-filter-parser-sql` module which can be used
with `EmbeddingStoreContentRetriever`'s `dynamicFilter` to automatically
build a `Filter` object from the `Query` using language model and
`SqlFilterParser`.
For example:
```java
TextSegment groundhogDay = TextSegment.from("Groundhog Day", new Metadata().put("genre", "comedy").put("year", 1993));
TextSegment forrestGump = TextSegment.from("Forrest Gump", new Metadata().put("genre", "drama").put("year", 1994));
TextSegment dieHard = TextSegment.from("Die Hard", new Metadata().put("genre", "action").put("year", 1998));
// describe metadata keys as if they were columns in the SQL table
TableDefinition tableDefinition = TableDefinition.builder()
.name("movies")
.addColumn("genre", "VARCHAR", "one of [comedy, drama, action]")
.addColumn("year", "INT")
.build();
LanguageModelSqlFilterBuilder sqlFilterBuilder = new LanguageModelSqlFilterBuilder(model, tableDefinition);
ContentRetriever contentRetriever = EmbeddingStoreContentRetriever.builder()
.embeddingStore(embeddingStore)
.embeddingModel(embeddingModel)
.dynamicFilter(sqlFilterBuilder::build)
.build();
String answer = assistant.answer("Recommend me a good drama from 90s"); // Forrest Gump
```
## Which embedding store integrations will support `Filter`?
In the long run, all (provided the embedding store itself supports it).
In the first iteration, I aim to add support to just a few:
- `InMemoryEmbeddingStore`
- Elasticsearch
- Milvus
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
## Summary by CodeRabbit
- **New Features**
- Introduced filters for checking key's value existence in a collection
for improved data handling.
- **Enhancements**
- Updated `InMemoryEmbeddingStoreTest` to extend a different class for
improved testing coverage and added a new test method.
- **Refactor**
- Made minor formatting adjustments in the assertion block for better
readability.
- **Documentation**
- Updated class hierarchy information for clarity.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
So far, LangChain4j had only a simple (a.k.a., naive) RAG
implementation: a single `Retriever` was invoked on each interaction
with the LLM, and all retrieved `TextSegments` were appended to the end
of the `UserMessage`. This approach was very limiting.
This PR introduces support for much more advanced RAG use cases. The
design and mental model are inspired by [this
article](https://blog.langchain.dev/deconstructing-rag/) and [this
paper](https://arxiv.org/abs/2312.10997), making it advisable to read
the article.
This PR introduces a `RetrievalAugmentor` interface responsible for
augmenting a `UserMessage` with relevant content before sending it to
the LLM. The `RetrievalAugmentor` can be used with both `AiServices` and
`ConversationalRetrievalChain`, as well as stand-alone.
A default implementation of `RetrievalAugmentor`
(`DefaultRetrievalAugmentor`) is provided with the library and is
suggested as a good starting point. However, users are not limited to it
and can have more freedom with their own custom implementations.
`DefaultRetrievalAugmentor` decomposes the entire RAG flow into more
granular steps and base components:
- `QueryTransformer`
- `QueryRouter`
- `ContentRetriever` (the old `Retriever` is now deprecated)
- `ContentAggregator`
- `ContentInjector`
This modular design aims to separate concerns and simplify development,
testing, and evaluation. Most (if not all) currently known and proven
RAG techniques can be represented as one or multiple base components
listed above.
Here is how the decomposed RAG flow can be visualized:
This mental and software model aims to simplify the thinking, reasoning,
and implementation of advanced RAG flows.
Each base component listed above has a sensible and simple default
implementation configured in `DefaultRetrievalAugmentor` by default but
can be overridden by more sophisticated implementations (provided by the
library out-of-the-box) as well as custom ones. The list of
implementations is expected to grow over time as we discover new
techniques and implement existing proven ones.
This PR also introduces out-of-the-box support for the following proven
RAG techniques:
- Query expansion
- Query compression
- Query routing using LLM
- [Reciprocal Rank
Fusion](https://learn.microsoft.com/en-us/azure/search/hybrid-search-ranking)
- Re-ranking ([Cohere Rerank](https://docs.cohere.com/docs/reranking)
integration is coming in a [separate
PR](https://github.com/langchain4j/langchain4j/pull/539)).
Qwen series models are provided by Alibaba Cloud. They are much better
in Asia languages then other LLMs.
DashScope is a model service platform. Qwen models are its primary
supported models. But it also supports other series like LLaMA2, Dolly,
ChatGLM, BiLLa(based on LLaMA)...These may be integrated sometime in the
future.
- all-minilm-l6-v2
- all-minilm-l6-v2-q
- e5-small-v2
- e5-small-v2-q
The idea is to give users an option to embed documents/texts in the same
Java process without any external dependencies.
ONNX Runtime is used to run models inside JVM.
Each model resides in it's own maven module (inside the jar).
Have a parent pom that contains most/all common things for the
sub-projects.
Note that it is separate from the root aggregator pom: not mixing the
aggregator and the parents makes things slightly easier.
If this change makes it harder to do releases, there might be a
possibility to generate the effective poms for each subproject, but on
the other hand releasing everything should not be too problematic.
The title says it all. Relying on [this maven
plugin](https://github.com/chonton/license-maven-plugin) for it.
Note that this adds a separate build step because we need a more recent
JDK to run the needed plugin.
LangChain4j