allow generate responses without source documents

Author: recchris

backend/CHANGELOG.md

@@ -9,6 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- Added support for queries without source data in vector database
 - Graceful failure of triple export when no chunks are found
 
 ## [v0.1.5] - 2024-10-29

backend/src/app/api/v1/endpoints/query.py

@@ -16,6 +16,7 @@
 from app.services.query_service import (
     decomposition_query,
     hybrid_query,
+    inference_query,
     simple_vector_query,
 )
 from app.services.vector_db.base import VectorDBService
@@ -59,6 +60,30 @@ async def run_query(
     HTTPException
         If there's an error processing the query.
     """
+    if request.document_id == "ffffffffffffffffffffffffffffffff":
+        query_response = await inference_query(
+            request.prompt.query,
+            request.prompt.rules,
+            request.prompt.type,
+            llm_service,
+        )
+
+        if not isinstance(query_response, QueryResult):
+            query_response = QueryResult(**query_response)
+
+        answer = QueryAnswer(
+            id=uuid.uuid4().hex,
+            document_id=request.document_id,
+            prompt_id=request.prompt.id,
+            answer=query_response.answer,
+            type=request.prompt.type,
+        )
+        response_data = QueryAnswerResponse(
+            answer=answer, chunks=query_response.chunks
+        )
+
+        return response_data
+
     try:
         logger.info(f"Received query request: {request.model_dump()}")
 

backend/src/app/models/llm_responses.py

@@ -82,6 +82,8 @@ def validate_array(
         v = cls.validate_none(v)
         if v is None:
             return None
+        if len(v) == 1 and v[0] == "None":
+            return None
         if not isinstance(v, list):
             raise ValueError("Must be a list or None")
         if max_length and len(v) > max_length:

backend/src/app/services/llm/prompts.py

@@ -24,17 +24,29 @@
 - Provide your answer based strictly on the given context.
 - Be concise and accurate.
 - Do not include any introductory or concluding remarks.
-- If the answer is not present in the context, respond exactly with `"None"` (without quotes).
+- If the answer is not present in the context, respond exactly with "None".
 
 **Answer**:
 """
 )
 
+INFERRED_BASE_PROMPT = Template(
+    """
+Answer the following question following the formatting instructions at the bottom. Do not include, quotes, formatting, or any explanation or extra information. Just answer the question.
+
+**Question**: $query
+**Answer**:
+
+$format_specific_instructions
+
+"""
+)
+
 BOOL_INSTRUCTIONS = """
 **Special Instructions for Boolean Questions**:
 
-- If the question is asking for a verification or requires a boolean answer, respond with `"True"` or `"False"` (as a string).
-- If you cannot determine the answer from the context, respond exactly with `"None"`.
+- If the question is asking for a verification or requires a boolean answer, respond with True or False.
+- If you cannot answer the question, respond exactly with 'None'.
 - Do not provide any explanations or additional information.
 """
 
@@ -45,9 +57,9 @@
 
 **Special Instructions for String Responses**:
 
-- If the answer is a single string, provide the string enclosed in double quotes.
+- If the answer is a single string, provide a single string.
 - If multiple strings are required, provide them as a JSON array of strings.
-- If you cannot find an answer, respond exactly with `"None"`.
+- If you cannot find an answer, respond exactly with 'None'.
 - Do not include any additional text or explanation.
 """
 )
@@ -60,7 +72,7 @@
 
 - If the answer is a single integer, provide the integer as a number.
 - If multiple integers are required, provide them as a JSON array of integers.
-- If you cannot find an answer, respond exactly with `"None"`.
+- If you cannot find an answer, respond exactly with 'None'.
 - Do not include any additional text or explanation.
 """
 )
@@ -79,7 +91,7 @@
 
 - Provide the keywords as a JSON array of strings.
 - Ensure all words are in their base (lemmatized) form.
-- If you cannot extract any relevant keywords, respond exactly with `"None"`.
+- If you cannot extract any relevant keywords, respond exactly with 'None'.
 - Do not include any additional text or explanation.
 
 **Keywords**:
@@ -105,7 +117,7 @@
 
 - Provide the similar keywords as a JSON array of strings.
 - Only include words that are present in the context and are semantically related to the provided keywords.
-- If you cannot find any similar keywords in the context, respond exactly with `"None"`.
+- If you cannot find any similar keywords in the context, respond exactly with 'None'.
 - Do not include any additional text or explanation.
 
 **Similar Keywords**:
@@ -125,7 +137,7 @@
 **Instructions**:
 
 - Provide up to 3 sub-questions as a JSON array of strings.
-- If the question is already simple or cannot be decomposed, respond exactly with `"None"`.
+- If the question is already simple or cannot be decomposed, respond exactly with 'None'.
 - Do not include any additional text or explanation.
 
 **Sub-Questions**:

backend/src/app/services/llm_service.py

@@ -21,6 +21,7 @@
     BASE_PROMPT,
     BOOL_INSTRUCTIONS,
     DECOMPOSE_QUERY_PROMPT,
+    INFERRED_BASE_PROMPT,
     INT_ARRAY_INSTRUCTIONS,
     KEYWORD_PROMPT,
     SCHEMA_PROMPT,
@@ -153,6 +154,58 @@ async def generate_response(
         return {"answer": None}
 
 
+async def generate_inferred_response(
+    llm_service: LLMService,
+    query: str,
+    rules: list[Rule],
+    format: FormatType,
+) -> dict[str, Any]:
+    """
+    Generate a response from the language model based on the given query and format.
+
+    Parameters
+    ----------
+    llm_service : LLMService
+        The language model service to use for generating the response.
+    query : str
+        The user's query to be answered.
+    rules : list[Rule]
+        A list of rules to apply when generating the response.
+    format : Literal["int", "str", "bool", "int_array", "str_array"]
+        The desired format of the response.
+
+    Returns
+    -------
+    dict[str, Any]
+        A dictionary containing the generated answer or None if an error occurs.
+    """
+    logger.info(
+        f"Generating inferred response for query: {query} in format: {format}"
+    )
+
+    output_model, format_specific_instructions = _get_model_and_instructions(
+        format, rules, query
+    )
+    prompt = INFERRED_BASE_PROMPT.substitute(
+        query=query,
+        format_specific_instructions=format_specific_instructions,
+    )
+
+    try:
+        response = await llm_service.generate_completion(prompt, output_model)
+        logger.info(f"Raw response from LLM: {response}")
+
+        if response is None or response.answer is None:
+            logger.warning("LLM returned None response")
+            return {"answer": None}
+
+        logger.info(f"Processed response: {response.answer}")
+        return {"answer": response.answer}
+    except Exception as e:
+        logger.error(f"Error generating response: {str(e)}", exc_info=True)
+        return {"answer": None}
+
+
 async def get_keywords(
     llm_service: LLMService, query: str
 ) -> dict[str, list[str] | None]:

backend/src/app/services/query_service.py

@@ -5,7 +5,11 @@
 
 from app.models.query_core import Chunk, FormatType, QueryType, Rule
 from app.schemas.query_api import QueryResult, SearchResponse
-from app.services.llm_service import LLMService, generate_response
+from app.services.llm_service import (
+    LLMService,
+    generate_inferred_response,
+    generate_response,
+)
 
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
@@ -124,3 +128,22 @@ async def simple_vector_query(
         llm_service,
         vector_db_service,
     )
+
+
+async def inference_query(
+    query: str,
+    rules: List[Rule],
+    format: FormatType,
+    llm_service: LLMService,
+) -> QueryResult:
+    """Generate a response, no need for vector retrieval."""
+
+    # Since we are just answering this query based on data provided in the query,
+    # ther is no need to retrieve any chunks from the vector database.
+
+    answer = await generate_inferred_response(
+        llm_service, query, rules, format
+    )
+    answer_value = answer["answer"]
+
+    return QueryResult(answer=answer_value, chunks=[])

frontend/CHANGELOG.md

@@ -9,6 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Changed
 
+- Allow query without source data if mention is present
 - Removed replace rule as an option for now
 
 ## [v0.1.5] - 2024-10-29

frontend/src/config/api.ts

@@ -71,7 +71,12 @@ export async function runQuery(
   column: AnswerTableColumn,
   globalRules: AnswerTableGlobalRule[]
 ) {
-  if (!row.sourceData || !column.entityType.trim() || !column.generate) {
+  // if (!row.sourceData || !column.entityType.trim() || !column.generate) {
+  //   throw new Error(
+  //     "Row or column doesn't allow running query (missing row source data or column is empty or has generate set to false)"
+  //   );
+  // }
+  if (!column.entityType.trim() || !column.generate) {
     throw new Error(
       "Row or column doesn't allow running query (missing row source data or column is empty or has generate set to false)"
     );
@@ -88,7 +93,9 @@ export async function runQuery(
       "Content-Type": "application/json"
     },
     body: JSON.stringify({
-      document_id: row.sourceData.document.id,
+      document_id: row.sourceData?.document?.id
+        ? row.sourceData.document.id
+        : "ffffffffffffffffffffffffffffffff",
       prompt: {
         id: column.id,
         entity_type: column.entityType,

frontend/src/config/store/store.ts

@@ -308,7 +308,6 @@ export const useStore = create<Store>()(
               row &&
               column.entityType.trim() &&
               column.generate &&
-              row.sourceData &&
               !loadingCells[key]
               ? { key, column, row }
               : null;
@@ -324,28 +323,42 @@ export const useStore = create<Store>()(
 
         for (const { key, row, column: column_ } of batch) {
           const column = cloneDeep(column_);
+          let shouldRunQuery = true;
+          let hasMatches = false;
 
           // Replace all column references with the row's answer to that column
           for (const [match, columnId] of column.query.matchAll(
             /@\[[^\]]+\]\(([^)]+)\)/g
           )) {
+            hasMatches = true;
             const targetColumn = columns.find(c => c.id === columnId);
             if (!targetColumn) continue;
             const cell = row.cells[targetColumn.id];
-            if (isNil(cell)) continue;
+            if (isNil(cell) || (isNil(cell) && isNil(row.sourceData))) {
+              shouldRunQuery = false;
+              break;
+            }
             column.query = column.query.replace(match, String(cell));
           }
-
-          runQuery(row, column, globalRules).then(({ answer, chunks }) => {
-            editCells(
-              [{ rowId: row.id, columnId: column.id, cell: answer.answer }],
-              activeTableId
-            );
+          if (!hasMatches && isNil(row.sourceData)) {
+            shouldRunQuery = false;
+          }
+          if (shouldRunQuery) {
+            runQuery(row, column, globalRules).then(({ answer, chunks }) => {
+              editCells(
+                [{ rowId: row.id, columnId: column.id, cell: answer.answer }],
+                activeTableId
+              );
+              editTable(activeTableId, {
+                chunks: { ...getTable(activeTableId).chunks, [key]: chunks },
+                loadingCells: omit(getTable(activeTableId).loadingCells, key)
+              });
+            });
+          } else {
             editTable(activeTableId, {
-              chunks: { ...getTable(activeTableId).chunks, [key]: chunks },
               loadingCells: omit(getTable(activeTableId).loadingCells, key)
             });
-          });
+          }
         }
       },