Search Regression Test and Save/Load State updates (#761)

backend/danswer/direct_qa/answer_question.py

@@ -248,7 +248,9 @@ def answer_qa_query_stream(
         batch_offset=offset_count,
     )
     llm_relevance_filtering_response = LLMRelevanceFilterResponse(
-        relevant_chunk_indices=llm_chunks_indices
+        relevant_chunk_indices=[
+            index for index, value in enumerate(llm_chunk_selection) if value
+        ]
     ).dict()
     yield get_json_line(llm_relevance_filtering_response)
 

backend/danswer/direct_qa/qa_utils.py

@@ -322,7 +322,7 @@ def get_usable_chunks(
 def get_chunks_for_qa(
     chunks: list[InferenceChunk],
     llm_chunk_selection: list[bool],
-    token_limit: int = NUM_DOCUMENT_TOKENS_FED_TO_GENERATIVE_MODEL,
+    token_limit: int | None = NUM_DOCUMENT_TOKENS_FED_TO_GENERATIVE_MODEL,
     batch_offset: int = 0,
 ) -> list[int]:
     """
@@ -353,13 +353,17 @@ def get_chunks_for_qa(
             token_count += chunk_token + 50
 
             # Always use at least 1 chunk
-            if token_count <= token_limit or not latest_batch_indices:
+            if (
+                token_limit is None
+                or token_count <= token_limit
+                or not latest_batch_indices
+            ):
                 latest_batch_indices.append(ind)
                 current_chunk_unused = False
             else:
                 current_chunk_unused = True
 
-            if token_count >= token_limit:
+            if token_limit is not None and token_count >= token_limit:
                 if batch_index < batch_offset:
                     batch_index += 1
                     if current_chunk_unused:

backend/danswer/search/search_runner.py

@@ -540,7 +540,7 @@ def full_chunk_search_generator(
         if llm_filter_task_id
         else None,
     )
-    if llm_chunk_selection:
+    if llm_chunk_selection is not None:
         yield [chunk.unique_id in llm_chunk_selection for chunk in retrieved_chunks]
     else:
         yield [True for _ in reranked_chunks or retrieved_chunks]

backend/scripts/save_load_state.py

@@ -1,4 +1,5 @@
 # This file is purely for development use, not included in any builds
+# Remember to first to send over the schema information (run API Server)
 import argparse
 import json
 import os
@@ -19,25 +20,40 @@ from danswer.utils.logger import setup_logger
 logger = setup_logger()
 
 
-def save_postgres(filename: str) -> None:
+def save_postgres(filename: str, container_name: str) -> None:
     logger.info("Attempting to take Postgres snapshot")
-    cmd = f"pg_dump -U {POSTGRES_USER} -h {POSTGRES_HOST} -p {POSTGRES_PORT} -W -F t {POSTGRES_DB} > {filename}"
-    subprocess.run(
-        cmd, shell=True, check=True, input=f"{POSTGRES_PASSWORD}\n", text=True
-    )
+    cmd = f"docker exec {container_name} pg_dump -U {POSTGRES_USER} -h {POSTGRES_HOST} -p {POSTGRES_PORT} -W -F t {POSTGRES_DB}"
+    with open(filename, "w") as file:
+        subprocess.run(
+            cmd,
+            shell=True,
+            check=True,
+            stdout=file,
+            text=True,
+            input=f"{POSTGRES_PASSWORD}\n",
+        )
 
 
-def load_postgres(filename: str) -> None:
+def load_postgres(filename: str, container_name: str) -> None:
     logger.info("Attempting to load Postgres snapshot")
     try:
         alembic_cfg = Config("alembic.ini")
         command.upgrade(alembic_cfg, "head")
-    except Exception:
-        logger.info("Alembic upgrade failed, maybe already has run")
-    cmd = f"pg_restore --clean -U {POSTGRES_USER} -h {POSTGRES_HOST} -p {POSTGRES_PORT} -W -d {POSTGRES_DB} -1 {filename}"
-    subprocess.run(
-        cmd, shell=True, check=True, input=f"{POSTGRES_PASSWORD}\n", text=True
+    except Exception as e:
+        logger.error(f"Alembic upgrade failed: {e}")
+
+    host_file_path = os.path.abspath(filename)
+
+    copy_cmd = f"docker cp {host_file_path} {container_name}:/tmp/"
+    subprocess.run(copy_cmd, shell=True, check=True)
+
+    container_file_path = f"/tmp/{os.path.basename(filename)}"
+
+    restore_cmd = (
+        f"docker exec {container_name} pg_restore --clean -U {POSTGRES_USER} "
+        f"-h localhost -p {POSTGRES_PORT} -d {POSTGRES_DB} -1 -F t {container_file_path}"
     )
+    subprocess.run(restore_cmd, shell=True, check=True)
 
 
 def save_vespa(filename: str) -> None:
@@ -85,6 +101,12 @@ if __name__ == "__main__":
     parser.add_argument(
         "--load", action="store_true", help="Load Danswer state from save directory."
     )
+    parser.add_argument(
+        "--postgres_container_name",
+        type=str,
+        default="danswer-stack-relational_db-1",
+        help="Name of the postgres container to dump",
+    )
     parser.add_argument(
         "--checkpoint_dir",
         type=str,
@@ -94,6 +116,7 @@ if __name__ == "__main__":
 
     args = parser.parse_args()
     checkpoint_dir = args.checkpoint_dir
+    postgres_container = args.postgres_container_name
 
     if not os.path.exists(checkpoint_dir):
         os.makedirs(checkpoint_dir)
@@ -102,8 +125,12 @@ if __name__ == "__main__":
         raise ValueError("Must specify --save or --load")
 
     if args.load:
-        load_postgres(os.path.join(checkpoint_dir, "postgres_snapshot.tar"))
+        load_postgres(
+            os.path.join(checkpoint_dir, "postgres_snapshot.tar"), postgres_container
+        )
         load_vespa(os.path.join(checkpoint_dir, "vespa_snapshot.jsonl"))
     else:
-        save_postgres(os.path.join(checkpoint_dir, "postgres_snapshot.tar"))
+        save_postgres(
+            os.path.join(checkpoint_dir, "postgres_snapshot.tar"), postgres_container
+        )
         save_vespa(os.path.join(checkpoint_dir, "vespa_snapshot.jsonl"))

backend/tests/regression/search_quality/eval_search.py

@@ -8,6 +8,7 @@ from typing import TextIO
 from sqlalchemy.orm import Session
 
 from danswer.db.engine import get_sqlalchemy_engine
+from danswer.direct_qa.qa_utils import get_chunks_for_qa
 from danswer.document_index.factory import get_default_document_index
 from danswer.indexing.models import InferenceChunk
 from danswer.search.models import IndexFilters
@@ -70,7 +71,6 @@ def get_search_results(
     query: str, db_session: Session
 ) -> tuple[
     list[InferenceChunk],
-    list[bool],
     RetrievalMetricsContainer | None,
     RerankMetricsContainer | None,
 ]:
@@ -89,7 +89,7 @@ def get_search_results(
     retrieval_metrics = MetricsHander[RetrievalMetricsContainer]()
     rerank_metrics = MetricsHander[RerankMetricsContainer]()
 
-    chunks, llm_filter, query_id = danswer_search(
+    top_chunks, llm_chunk_selection, query_id = danswer_search(
         question=question,
         user=None,
         db_session=db_session,
@@ -99,16 +99,23 @@ def get_search_results(
         rerank_metrics_callback=rerank_metrics.record_metric,
     )
 
+    llm_chunks_indices = get_chunks_for_qa(
+        chunks=top_chunks,
+        llm_chunk_selection=llm_chunk_selection,
+        token_limit=None,
+    )
+
+    llm_chunks = [top_chunks[i] for i in llm_chunks_indices]
+
     return (
-        chunks,
-        llm_filter,
+        llm_chunks,
         retrieval_metrics.metrics,
         rerank_metrics.metrics,
     )
 
 
 def _print_retrieval_metrics(
-    metrics_container: RetrievalMetricsContainer, show_all: bool
+    metrics_container: RetrievalMetricsContainer, show_all: bool = False
 ) -> None:
     for ind, metric in enumerate(metrics_container.metrics):
         if not show_all and ind >= 10:
@@ -117,14 +124,13 @@ def _print_retrieval_metrics(
         if ind != 0:
             print()  # for spacing purposes
         print(f"\tDocument: {metric.document_id}")
-        print(f"\tLink: {metric.first_link or 'NA'}")
         section_start = metric.chunk_content_start.replace("\n", " ")
         print(f"\tSection Start: {section_start}")
         print(f"\tSimilarity Distance Metric: {metric.score}")
 
 
 def _print_reranking_metrics(
-    metrics_container: RerankMetricsContainer, show_all: bool
+    metrics_container: RerankMetricsContainer, show_all: bool = False
 ) -> None:
     # Printing the raw scores as they're more informational than post-norm/boosting
     for ind, metric in enumerate(metrics_container.metrics):
@@ -134,45 +140,81 @@ def _print_reranking_metrics(
         if ind != 0:
             print()  # for spacing purposes
         print(f"\tDocument: {metric.document_id}")
-        print(f"\tLink: {metric.first_link or 'NA'}")
         section_start = metric.chunk_content_start.replace("\n", " ")
         print(f"\tSection Start: {section_start}")
         print(f"\tSimilarity Score: {metrics_container.raw_similarity_scores[ind]}")
 
 
-def main(questions_json: str, output_file: str) -> None:
+def calculate_score(
+    log_prefix: str, chunk_ids: list[str], targets: list[str], max_chunks: int = 5
+) -> float:
+    top_ids = chunk_ids[:max_chunks]
+    matches = [top_id for top_id in top_ids if top_id in targets]
+    print(f"{log_prefix} Hits: {len(matches)}/{len(targets)}", end="\t")
+    return len(matches) / min(len(targets), max_chunks)
+
+
+def main(
+    questions_json: str, output_file: str, show_details: bool, stop_after: int
+) -> None:
     questions_info = read_json(questions_json)
 
+    running_retrieval_score = 0.0
+    running_rerank_score = 0.0
+    running_llm_filter_score = 0.0
+
     with open(output_file, "w") as outfile:
         with redirect_print_to_file(outfile):
             print("Running Document Retrieval Test\n")
 
             with Session(engine, expire_on_commit=False) as db_session:
-                for question, targets in questions_info.items():
-                    print(f"Question: {question}")
+                for ind, (question, targets) in enumerate(questions_info.items()):
+                    if ind >= stop_after:
+                        break
+
+                    print(f"\n\nQuestion: {question}")
 
                     (
-                        chunks,
-                        llm_filters,
+                        top_chunks,
                         retrieval_metrics,
                         rerank_metrics,
                     ) = get_search_results(query=question, db_session=db_session)
 
-                    print("\nRetrieval Metrics:")
-                    if retrieval_metrics is None:
-                        print("No Retrieval Metrics Available")
-                    else:
-                        _print_retrieval_metrics(
-                            retrieval_metrics, show_all=args.all_results
-                        )
+                    assert retrieval_metrics is not None and rerank_metrics is not None
 
-                    print("\nReranking Metrics:")
-                    if rerank_metrics is None:
-                        print("No Reranking Metrics Available")
-                    else:
-                        _print_reranking_metrics(
-                            rerank_metrics, show_all=args.all_results
-                        )
+                    retrieval_ids = [
+                        metric.document_id for metric in retrieval_metrics.metrics
+                    ]
+                    retrieval_score = calculate_score(
+                        "Retrieval", retrieval_ids, targets
+                    )
+                    running_retrieval_score += retrieval_score
+                    print(f"Average: {running_retrieval_score / (ind + 1)}")
+
+                    rerank_ids = [
+                        metric.document_id for metric in rerank_metrics.metrics
+                    ]
+                    rerank_score = calculate_score("Rerank", rerank_ids, targets)
+                    running_rerank_score += rerank_score
+                    print(f"Average: {running_rerank_score / (ind + 1)}")
+
+                    llm_ids = [chunk.document_id for chunk in top_chunks]
+                    llm_score = calculate_score("LLM Filter", llm_ids, targets)
+                    running_llm_filter_score += llm_score
+                    print(f"Average: {running_llm_filter_score / (ind + 1)}")
+
+                    if show_details:
+                        print("\nRetrieval Metrics:")
+                        if retrieval_metrics is None:
+                            print("No Retrieval Metrics Available")
+                        else:
+                            _print_retrieval_metrics(retrieval_metrics)
+
+                        print("\nReranking Metrics:")
+                        if rerank_metrics is None:
+                            print("No Reranking Metrics Available")
+                        else:
+                            _print_reranking_metrics(rerank_metrics)
 
 
 if __name__ == "__main__":
@@ -190,6 +232,23 @@ if __name__ == "__main__":
         help="Path to the output results file.",
         default="./tests/regression/search_quality/regression_results.txt",
     )
+    parser.add_argument(
+        "--show_details",
+        action="store_true",
+        help="If set, show details of the retrieved chunks.",
+        default=True,
+    )
+    parser.add_argument(
+        "--stop_after",
+        type=int,
+        help="Stop processing after this many iterations.",
+        default=10,
+    )
     args = parser.parse_args()
 
-    main(args.regression_questions_json, args.output_file)
+    main(
+        args.regression_questions_json,
+        args.output_file,
+        args.show_details,
+        args.stop_after,
+    )