Refactoring Deckretriever (#7671)

Co-authored-by: Kamil Piechowiak <[email protected]>
GitOrigin-RevId: 0212b7d459cf1767ed68635aa42fff179c435fd4

CHANGELOG.md

@@ -6,12 +6,14 @@ This project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 ## [Unreleased]
 
 ### Added
+- `pw.xpacks.llm.document_store.SlidesDocumentStore`, which is a subclass of `pw.xpacks.llm.document_store.DocumentStore` customized for retrieving slides from presentations.
 - `pw.temporal.inactivity_detection` and `pw.temporal.utc_now` functions allowing for alerting and other time dependent usecases
 
 ### Changed
 - `pw.Table.concat`, `pw.Table.with_id`, `pw.Table.with_id_from` no longer perform checks if ids are unique. It improves memory usage.
 - table operations that store values (like `pw.Table.join`, `pw.Table.update_cells`) no longer store columns that are not used downstream.
 - `append_only` column property is now propagated better (there are more places where we can infer it).
+- **BREAKING**: Unused arguments from the constructor `pw.xpacks.llm.question_answering.DeckRetriever` are no longer accepted. 
 
 ### Fixed
 - `query_as_of_now` of `pw.stdlib.indexing.DataIndex` and `pw.stdlib.indexing.HybridIndex` now work in constant memory for infinite query stream (no query-related data is kept after query is answered).

python/pathway/xpacks/llm/document_store.py

@@ -8,7 +8,7 @@
 """
 
 from collections.abc import Callable
-from typing import TYPE_CHECKING, Iterable
+from typing import TYPE_CHECKING, Iterable, TypeAlias
 
 import jmespath
 
@@ -185,7 +185,7 @@ class FilterSchema(pw.Schema):
             default_value=None, description="An optional Glob pattern for the file path"
         )
 
-    InputsQuerySchema = FilterSchema
+    InputsQuerySchema: TypeAlias = FilterSchema
 
     class InputsResultSchema(pw.Schema):
         result: list[pw.Json]
@@ -365,7 +365,7 @@ def _get_jmespath_filter(
 
     @pw.table_transformer
     def inputs_query(
-        self, input_queries: pw.Table[InputsQuerySchema]  # type: ignore
+        self, input_queries: pw.Table[InputsQuerySchema]
     ) -> pw.Table[InputsResultSchema]:
         """
         Query ``DocumentStore`` for the list of input documents.
@@ -448,3 +448,64 @@ def retrieve_query(
     @property
     def index(self) -> DataIndex:
         return self._retriever
+
+
+class SlidesDocumentStore(DocumentStore):
+    """
+    Document store for the ``slide-search`` application.
+    Builds a document indexing pipeline and starts an HTTP REST server.
+
+    Adds to the ``DocumentStore`` a new method ``parsed_documents`` a set of
+    documents metadata after the parsing and document post processing stages.
+    """
+
+    excluded_response_metadata = ["b64_image"]
+
+    @pw.table_transformer
+    def parsed_documents_query(
+        self,
+        parse_docs_queries: pw.Table[DocumentStore.InputsQuerySchema],
+    ) -> pw.Table:
+        """
+        Query the SlidesDocumentStore for the list of documents with the associated
+        metadata after the parsing stage.
+        """
+        docs = self.parsed_docs
+
+        all_metas = docs.reduce(metadatas=pw.reducers.tuple(pw.this.metadata))
+
+        parse_docs_queries = self.merge_filters(parse_docs_queries)
+
+        @pw.udf
+        def format_inputs(
+            metadatas: list[pw.Json] | None,
+            metadata_filter: str | None,
+        ) -> list[pw.Json]:
+            metadatas = metadatas if metadatas is not None else []
+            if metadata_filter:
+                metadatas = [
+                    m
+                    for m in metadatas
+                    if jmespath.search(
+                        metadata_filter, m.value, options=_knn_lsh._glob_options
+                    )
+                ]
+
+            metadata_list: list[dict] = [m.as_dict() for m in metadatas]
+
+            for metadata in metadata_list:
+                for metadata_key in self.excluded_response_metadata:
+                    metadata.pop(metadata_key, None)
+
+            return [pw.Json(m) for m in metadata_list]
+
+        input_results = parse_docs_queries.join_left(
+            all_metas, id=parse_docs_queries.id
+        ).select(
+            all_metas.metadatas,
+            parse_docs_queries.metadata_filter,
+        )
+        input_results = input_results.select(
+            result=format_inputs(pw.this.metadatas, pw.this.metadata_filter)
+        )
+        return input_results

python/pathway/xpacks/llm/question_answering.py

@@ -11,13 +11,17 @@
 from pathway.internals import ColumnReference, Table, udfs
 from pathway.stdlib.indexing import DataIndex
 from pathway.xpacks.llm import Doc, llms, prompts
-from pathway.xpacks.llm.document_store import DocumentStore
+from pathway.xpacks.llm.document_store import DocumentStore, SlidesDocumentStore
 from pathway.xpacks.llm.llms import BaseChat, prompt_chat_single_qa
 from pathway.xpacks.llm.prompts import prompt_qa_geometric_rag
-from pathway.xpacks.llm.vector_store import VectorStoreClient, VectorStoreServer
+from pathway.xpacks.llm.vector_store import (
+    SlidesVectorStoreServer,
+    VectorStoreClient,
+    VectorStoreServer,
+)
 
 if TYPE_CHECKING:
-    from pathway.xpacks.llm.servers import QASummaryRestServer
+    from pathway.xpacks.llm.servers import QARestServer, QASummaryRestServer
 
 
 @pw.udf
@@ -455,14 +459,23 @@ def summarize_query(self, summarize_queries: pw.Table) -> pw.Table:
 
     @pw.table_transformer
     def retrieve(self, retrieve_queries: pw.Table) -> pw.Table:
+        """
+        Retrieve documents from the index.
+        """
         return self.indexer.retrieve_query(retrieve_queries)
 
     @pw.table_transformer
     def statistics(self, statistics_queries: pw.Table) -> pw.Table:
+        """
+        Get statistics about indexed files.
+        """
         return self.indexer.statistics_query(statistics_queries)
 
     @pw.table_transformer
     def list_documents(self, list_documents_queries: pw.Table) -> pw.Table:
+        """
+        Get list of documents from the retriever.
+        """
         return self.indexer.inputs_query(list_documents_queries)
 
     def build_server(
@@ -682,14 +695,45 @@ def answer_query(self, pw_ai_queries: pw.Table) -> pw.Table:
         return result
 
 
-class DeckRetriever(BaseRAGQuestionAnswerer):
-    """Class for slides search."""
+class DeckRetriever(BaseQuestionAnswerer):
+    """
+    Builds the logic for the Retriever of slides.
+
+    Args:
+        indexer: document store for parsing and indexing slides.
+        search_topk: Number of slides to be returned by the `answer_query` method.
+    """
 
     excluded_response_metadata = ["b64_image"]
 
+    def __init__(
+        self,
+        indexer: SlidesDocumentStore | SlidesVectorStoreServer,
+        *,
+        search_topk: int = 6,
+    ) -> None:
+        self.indexer = indexer
+        self._init_schemas()
+        self.search_topk = search_topk
+
+        self.server: None | QARestServer = None
+        self._pending_endpoints: list[tuple] = []
+
+    def _init_schemas(
+        self,
+    ) -> None:
+        class PWAIQuerySchema(pw.Schema):
+            prompt: str
+            filters: str | None = pw.column_definition(default_value=None)
+
+        self.AnswerQuerySchema = PWAIQuerySchema
+        self.RetrieveQuerySchema = self.indexer.RetrieveQuerySchema
+        self.StatisticsQuerySchema = self.indexer.StatisticsQuerySchema
+        self.InputsQuerySchema = self.indexer.InputsQuerySchema
+
     @pw.table_transformer
     def answer_query(self, pw_ai_queries: pw.Table) -> pw.Table:
-        """Return similar docs from the index."""
+        """Return slides similar to the given query."""
 
         pw_ai_results = pw_ai_queries + self.indexer.retrieve_query(
             pw_ai_queries.select(
@@ -720,6 +764,46 @@ def _format_results(docs: pw.Json) -> pw.Json:
 
         return pw_ai_results
 
+    @pw.table_transformer
+    def retrieve(self, retrieve_queries: pw.Table) -> pw.Table:
+        return self.indexer.retrieve_query(retrieve_queries)
+
+    @pw.table_transformer
+    def statistics(self, statistics_queries: pw.Table) -> pw.Table:
+        return self.indexer.statistics_query(statistics_queries)
+
+    @pw.table_transformer
+    def list_documents(self, list_documents_queries: pw.Table) -> pw.Table:
+        return self.indexer.parsed_documents_query(list_documents_queries)
+
+    def build_server(
+        self,
+        host: str,
+        port: int,
+        **rest_kwargs,
+    ):
+        warn(
+            "build_server method is deprecated. Instead, use explicitly a server from pw.xpacks.llm.servers.",
+            DeprecationWarning,
+            stacklevel=2,
+        )
+        # circular import
+        from pathway.xpacks.llm.servers import QARestServer
+
+        self.server = QARestServer(host, port, self, **rest_kwargs)
+
+    def run_server(self, *args, **kwargs):
+        warn(
+            "run_server method is deprecated. Instead, use explicitly a server from pw.xpacks.llm.servers.",
+            DeprecationWarning,
+            stacklevel=2,
+        )
+        if self.server is None:
+            raise ValueError(
+                "HTTP server is not built, initialize it with `build_server`"
+            )
+        self.server.run(*args, **kwargs)
+
 
 def send_post_request(
     url: str, data: dict, headers: dict = {}, timeout: int | None = None

python/pathway/xpacks/llm/vector_store.py

@@ -13,7 +13,7 @@
 import logging
 import threading
 from collections.abc import Callable, Coroutine
-from typing import TYPE_CHECKING, cast
+from typing import TYPE_CHECKING, TypeAlias, cast
 
 import jmespath
 import requests
@@ -330,7 +330,7 @@ class FilterSchema(pw.Schema):
             default_value=None, description="An optional Glob pattern for the file path"
         )
 
-    InputsQuerySchema = FilterSchema
+    InputsQuerySchema: TypeAlias = FilterSchema
 
     @staticmethod
     def merge_filters(queries: pw.Table):
@@ -363,7 +363,7 @@ def _get_jmespath_filter(
 
     @pw.table_transformer
     def inputs_query(
-        self, input_queries: pw.Table[InputsQuerySchema]  # type:ignore
+        self, input_queries: pw.Table[InputsQuerySchema]
     ) -> pw.Table[InputResultSchema]:
         docs = self._graph["docs"]
         # TODO: compare this approach to first joining queries to dicuments, then filtering,
@@ -576,7 +576,7 @@ class SlidesVectorStoreServer(VectorStoreServer):
     @pw.table_transformer
     def inputs_query(
         self,
-        input_queries: pw.Table[VectorStoreServer.InputsQuerySchema],  # type:ignore
+        input_queries: pw.Table[VectorStoreServer.InputsQuerySchema],
     ) -> pw.Table:
         docs = self._graph["parsed_docs"]
 
@@ -617,6 +617,13 @@ def format_inputs(
         )
         return input_results
 
+    @pw.table_transformer
+    def parsed_documents_query(
+        self,
+        parse_docs_queries: pw.Table[VectorStoreServer.InputsQuerySchema],
+    ) -> pw.Table:
+        return self.inputs_query(parse_docs_queries)
+
 
 class VectorStoreClient:
     """