Spaces:

10gen
/

deepsearchitv2

Runtime error

App Files Files Community

Guiyom commited on Feb 14, 2025

Commit

10fd028

verified ·

1 Parent(s): 93ae956

Update app.py

Browse files

Files changed (1) hide show

app.py +34 -41

app.py CHANGED Viewed

@@ -37,25 +37,11 @@ def call_visual_llm(prompt: str) -> str:
     return response
 def generate_visual_snippet(placeholder_text: str, context: str, initial_query: str, crumbs: str) -> str:
-    """
-    Given the placeholder instructions, create a prompt and call the LLM to generate
-    a complete HTML/CSS/JS snippet for the visualization.
-    """
     prompt = (f"""
-Generate a complete, self-contained HTML code snippet that includes inline CSS and JavaScript(Flexbox/Grid, animations, transitions).
-The code should display an interactive, appealing visualization based on the following requirements:
 {placeholder_text}
-// Requirements
-- the dimensions should be less than 400 high and 400 width
-- use a font no larger than 10, with bold and italic if needed
-- if for a specific shape the background is dark, the text should be white (and vice versa if the background is clear)
-- Use semantic HTML5 elements
-- Add subtle animations and transitions
-- Display either:
-o chart (histogram, curve) with the proper call to a js library
-o a diagram (in the style of a mindmap, or five forces, or a flow chart)
 // Reference
 The visual is expected to be integrated within a report generated for the user, it should make use of any relevant information from:
 - the initial user query:
@@ -65,12 +51,21 @@ The visual is expected to be integrated within a report generated for the user,
 - some knowledge material gathered from search engines
 {crumbs}
 // IMPORTTANT
-- The visualization should be responsive, include any necessary interactive features (such as tooltips, clickable items, animations),
-and output only the code
 - no extra explanation
-- no code fences.
 """
     )
     result = call_visual_llm(prompt)
@@ -171,10 +166,6 @@ def openai_call(prompt: str, messages: list = None, model: str = "o3-mini",
         return err_msg
 def analyze_with_gpt4o(query: str, snippet: str, breadth: int, temperature: float = 0.7, max_tokens: int = 4000) -> dict:
-    """
-    Use gpt-4o-mini to process a snippet from a query result.
-    Returns a dictionary with keys: 'relevant', 'summary', and 'followups'.
-    """
     client = openai.OpenAI(api_key=os.getenv('OPENAI_API_KEY'))
     prompt = (f"""Analyze the following content from a query result:
@@ -184,24 +175,26 @@ Research topic:
 {query}
 Instructions:
-1.  **Relevance:** Determine if the content is relevant to the research topic.  Answer with a single word: 'yes' or 'no'.
-2.  **Structured Summary (if relevant):** If the content is relevant, provide a comprehensive summary structured into the following sections.  **Prioritize extreme conciseness and token efficiency while preserving all key information.** Aim for the shortest possible summary that retains all essential facts, figures, arguments, and quotes. The total summary should not exceed 1000 words, but shorter is strongly preferred.
-    *   **Key Facts:**  List the core factual claims. Use short, declarative sentences or bullet points. **Apply lemmatization, common abbreviations (e.g., vs., e.g., i.e., AI, LLM), and remove unnecessary words.**
-    *   **Key Figures:** Extract numerical data, statistics, dates, percentages. Use numerical representation. **Present concisely (list or table format).**
-    *   **Key Arguments:** Identify main arguments/claims. Summarize supporting evidence and counter-arguments. **Use lemmatization, abbreviations, and concise phrasing. Remove redundant phrases.**
-    *   **Key Quotes:** Include significant quotes. Attribute quotes correctly. **Choose quotes that are concise and impactful. If a quote can be paraphrased concisely without losing essential meaning, paraphrase it and note that it's a paraphrase.** Use symbols instead of words (&, +, ->, =, ...).
-   **General Optimization Guidelines:**
-    *   **Lemmatize:** Use the root form of words (e.g., "running" -> "run").
-    *   **Abbreviate:** Use common abbreviations (see list above).
-    *   **Remove Redundancy:** Eliminate unnecessary words and phrases. Be concise.
-    *   **Shorten Words (Carefully):** If a shorter word conveys the same meaning (e.g., "information" -> "info"), use it, but avoid ambiguity.
-    * **Implicit Representation:** Remove redundant terms.
-    * **Use Symbols:** Use symbols instead of words (&, +, ->, =, ...).
-3.  **Follow-up Search Queries:** Generate at least {breadth} follow-up search queries. These should be relevant to the research topic but also developments from the content summarized, aim for deeper understanding, use search operators (AND, OR, quotation marks), and be represented as a Python list of strings.
 For example: "Artificial intelligence" AND (mathematics OR geometry) -algebra,science AND history AND mathematics,...
-Return the result as a JSON object with the keys 'relevant', 'summary', and 'followups'. The 'summary' value should itself be a JSON object with keys 'Key Facts', 'Key Figures', 'Key Arguments', and 'Key Quotes'.
 Proceed."""
     )

     return response
 def generate_visual_snippet(placeholder_text: str, context: str, initial_query: str, crumbs: str) -> str:
     prompt = (f"""
+Generate a complete, self-contained HTML code snippet that includes inline CSS and JavaScript (only to call relevant libraries).
+The code should display a simple but effective and elegant visualization based on the following requirements:
 {placeholder_text}
 // Reference
 The visual is expected to be integrated within a report generated for the user, it should make use of any relevant information from:
 - the initial user query:
 - some knowledge material gathered from search engines
 {crumbs}
+// Requirements
+- the dimensions should be less than 500px height and 500px width (it should be printable once the report is converted to pdf)
+- use a font no larger than 10, with bold and italic if needed
+- if for a specific shape the background is dark, the text should be white (and vice versa if the background is clear)
+- Use HTML5 elements if necessary
+- Display either:
+o chart (histogram, curve) with the proper call to a js library (ex: d3.js or plotly)
+o a diagram (in the style of a mindmap, or five forces, or a flow chart)
+- keep it simple but effective to convey the message
 // IMPORTTANT
+- output only the code
 - no extra explanation
+- no code fences
+- do not add <html> </html> or  <!DOCTYPE html>, the snippet will be integrated in a html code body part at a pre-defined location
 """
     )
     result = call_visual_llm(prompt)
         return err_msg
 def analyze_with_gpt4o(query: str, snippet: str, breadth: int, temperature: float = 0.7, max_tokens: int = 4000) -> dict:
     client = openai.OpenAI(api_key=os.getenv('OPENAI_API_KEY'))
     prompt = (f"""Analyze the following content from a query result:
 {query}
 Instructions:
+1.  Relevance: Determine if the content is relevant to the research topic.  Answer with a single word: 'yes' or 'no'.
+2.  Structure: If the content is relevant, provide a comprehensive summary structured into the following sections.  Prioritize extreme conciseness and token efficiency while preserving all key information. Aim for the shortest possible summary that retains all essential facts, figures, arguments, and quotes. The total summary should not exceed 1000 words, but shorter is strongly preferred.
+    -   Key Facts (at least 10):  List the core factual claims. Use short, declarative sentences or bullet points. Apply lemmatization, common abbreviations (e.g., vs., e.g., i.e., AI, LLM), and remove unnecessary words.
+    -   Key Figures (at least 5): Extract numerical data, statistics, dates, percentages. Use numerical representation. Present concisely (list or table format).
+    -   Key Arguments (at least 10): Identify main arguments/claims. Summarize supporting evidence and counter-arguments. Use lemmatization, abbreviations, and concise phrasing. Remove redundant phrases.
+    -   Key Quotes (at least 1 f any): Include significant quotes (with the name of the author between parenthesis). Attribute quotes correctly. Choose quotes that are concise and impactful. If a quote can be paraphrased concisely without losing essential meaning, paraphrase it and note that it's a paraphrase. Use symbols instead of words (&, +, ->, =, ...).
+    -   Structured summary (10 to 50 sentences depending on the length): mention anecdotes, people, locations, anything that make will make the end report relatable and grounded
+Note: General Optimization Guidelines:
+    -   Lemmatize: Use the root form of words (e.g., "running" -> "run").
+    -   Abbreviate: Use common abbreviations
+    -   Remove Redundancy: Eliminate unnecessary words and phrases. Be concise.
+    -   Shorten Words (Carefully): If a shorter word conveys the same meaning (e.g., "information" -> "info"), use it, but avoid ambiguity.
+    -   Implicit Representation: Remove redundant terms.
+    -    Use Symbols: Use symbols instead of words (&, +, ->, =, ...).
+3.  Follow-up Search Queries: Generate at least {breadth} follow-up search queries. These should be relevant to the research topic but also developments from the content summarized, aim for deeper understanding, use search operators (AND, OR, quotation marks), and be represented as a Python list of strings.
 For example: "Artificial intelligence" AND (mathematics OR geometry) -algebra,science AND history AND mathematics,...
+Return the result as a JSON object with the keys 'relevant', 'structure', and 'followups'. The 'structure' value should itself be a JSON object with keys 'Key Facts', 'Key Figures', 'Key Arguments', 'Key Quotes' and 'Summary'.
 Proceed."""
     )