fix: revert some changes to reranker prompt

Author: charnould

utils/rank-chunks.ts

@@ -232,16 +232,15 @@ ${context.query?.standalone_questions.length !== 0 ? context.query?.standalone_q
 
 # Evaluation Steps
 
-Think step by step and describe in 10-20 words your reasoning for choosing these scores.
-Tasks are totally independant. Today is ${today_is()}.
+- Think step by step and describe in 10-20 words your reasoning for choosing these scores. 
+- Tasks are totally independant.
+- Today is ${today_is()}.
 
 ## Task: Identify the Building of the Chunk
+
 ${
   context.query?.named_entities.building !== null
-    ? `
-  
-  Assess whether the chunk explicitly or implicitly refers to a building (e.g., building, house, residence, housing program, etc.) with a name resembling “${context.query?.named_entities.building}”. Be cautious of cases where the name might also belong to a person, such as distinguishing between a person named “Jean Racine” and a house called “Racine”:
-
+    ? `Assess whether the chunk explicitly or implicitly refers to a building (e.g., building, house, residence, housing program, etc.) with a name resembling “${context.query?.named_entities.building}”. Be cautious of cases where the name might also belong to a person, such as distinguishing between a person named “Jean Racine” and a house called “Racine”:
   - Score 0: The chunk mention another building.
 	- Score 1000: The chunk explicitly and unambiguously discusses the building in question.
 	- Score 1-999: The reference to a building is unclear, ambiguous, or only loosely connected, requiring further clarification.
@@ -253,10 +252,7 @@ ${
 
 ${
   context.query?.named_entities.process !== null
-    ? `
-
-Evaluate whether the chunk pertains to the process “${context.query?.named_entities.process}”, specifically focusing on the company’s procedures, workflows, or standard ways of working. This includes assessing if the chunk describes, references, or aligns with the operational or organizational methods associated with the specified process:
-
+    ? `Evaluate whether the chunk pertains to the process “${context.query?.named_entities.process}”, specifically focusing on the company’s procedures, workflows, or standard ways of working. This includes assessing if the chunk describes, references, or aligns with the operational or organizational methods associated with the specified process:
 - Score 0: The chunk mention another process or no process is evocated
 - Score 1000: The chunk explicitly addresses the process with high accuracy.
 - Score 1-999: Partial or nuanced relevance (e.g., related processes but not an exact match). Example: “Une panne d’ascenseur” is distinct from “Un locataire bloqué dans l’ascenseur”.
@@ -266,9 +262,10 @@ Evaluate whether the chunk pertains to the process “${context.query?.named_ent
 
 ## Task: Evaluate Overall Relevance
 
-Assess the chunk’s alignment with the user query, considering that relevant answers might be explicit, implicit, or require interpretation from the context. Keep in mind partial relevance is still highly valuable. Use the following criteria:
-1. Answer Precision
-  - Determines if the chunk addresses the query intent, even partially.
+Assess the chunk’s alignment with the user query, considering that relevant answers might be explicit, implicit, or require interpretation from the context. Use the following criteria:
+
+1. Direct Answer Precision
+  - Determines if the chunk directly addresses the query intent.
 	- Evaluates how clear, comprehensive, and unambiguous the response is.
 2. Semantic Matching
   - Examines the degree of alignment between the query’s intent and the chunk’s content.
@@ -278,11 +275,13 @@ Assess the chunk’s alignment with the user query, considering that relevant an
 	- Assesses how actionable, precise, and unambiguous the information is, even if the answer is partially hidden within broader context.
 
 Scoring Scale:
-- **1000 (Perfect Match)**: The chunk fully addresses the query, offers clear and comprehensive information, provides multiple confirmatory points, and includes actionable insights.
-- **500-999 (Partial Match)**: The chunk contains at least part of the answer to the query. The score should reflect the degree of alignment, detail, and clarity. **If the chunk contains any relevant part of the answer, it must score at least 500**.
-- **0-499 (Low or No Match)**: The chunk shows limited or no connection to the query. Scores in this range should be reserved for cases where relevance is unclear or missing entirely. Before assigning a score below 500, carefully re-check the chunk for overlooked relevance or implicit connections.
+- 1000 (Perfect Match): Fully answers the query, provides multiple confirmatory points, comprehensive explanation, and actionable information.
+- 1-999 (Partial Match): Varies based on the degree of alignment, detail, and clarity.
+- 0 (No Match): No relevant connection to the query.
+
+**For low scores (<500), re-check the chunk for **overlooked** relevance.**
 
-Please proceed with your analysis and evaluation of the given query and chunk
+Please proceed with your analysis and evaluation of the given query and chunk.
 
 `.trim()
   })