Spaces:

10gen
/

deepsearchitv2

Runtime error

App Files Files Community

Guiyom commited on Mar 1, 2025

Commit

354a2a1

verified ·

1 Parent(s): 40373c8

Update app.py

Browse files

Files changed (1) hide show

app.py +35 -7

app.py CHANGED Viewed

@@ -405,7 +405,7 @@ def generate_graph_snippet(placeholder_text: str, context: str, initial_query: s
         padding: 20px;
       }
       h2 {
-        text-align: center;
         margin-bottom: 0;
       }
       #chart {
@@ -534,7 +534,7 @@ def generate_graph_snippet(placeholder_text: str, context: str, initial_query: s
       font-family: Arial, sans-serif;
     }
     h1 {
-      text-align: center;
       margin: 20px 0;
     }
     #chart {
@@ -692,7 +692,7 @@ def generate_graph_snippet(placeholder_text: str, context: str, initial_query: s
       background-color: #f5f5f5;
     }
     h1 {
-      text-align: center;
       margin-bottom: 10px;
       color: #333;
     }
@@ -870,6 +870,7 @@ Keep in mind the:
 // Important
 - Make the visuals content rich, there's no point having a visual if its content has no real value.
 - It has to convey some relevant insights.
 - Take a deep breath, think step by step and think it well.
 - Use your judgement to decide between box plots, bubble charts, calendar view, chord diagrams, histograms, ...
 - Your response should start with <html> and end with </html> - no intro before, no comments after.
@@ -1217,6 +1218,18 @@ Note: no need to add a reference table at the end of the focus box, all the refe
 - Skip lines and add intermediate titles for key paragraphs
 - You have freedom on the structure, but it has to cover all potential aspects on the topic in between 500 and 1000 words
 // Important
 - Make it real, with anecdotes from the content
 - Ground the content on the inputs shared above (context and knowledge inputs) with facts and numbers
@@ -1227,6 +1240,7 @@ Note: no need to add a reference table at the end of the focus box, all the refe
     result = openai_call(prompt, model="o3-mini", max_tokens_param=10000)
     result = result.strip().strip("```").strip()
     logging.info(f"The code produced for this focus placeholder:\n{placeholder_text}\n\n {result}\n\n")
     return result
@@ -1466,7 +1480,7 @@ Instructions:
     -   Key Facts (at least 5): List the core factual claims using short, declarative sentences or bullet points. Apply lemmatization and standard abbreviations.
     -   Key Figures (at least 5): Extract numerical data (statistics, dates, percentages) and include any necessary context (units, references, explanations) required to interpret these numbers. Present them concisely (list or table format).
     -   Key Arguments (at least 5): Identify main arguments or claims. Summarize supporting evidence and counter-arguments concisely.
-    -   Key Quotes (at least 1 if any): Include significant quotes (with the author's name in parentheses). Attribute quotes correctly. Paraphrase if needed, indicating that it's a paraphrase. Use symbols (e.g., &, +, ->, =) to conserve tokens.
     -   Structured Summary (10 to 50 sentences): Provide a structured summary that includes anecdotes, people, and locations to ensure the report is relatable.
 Note: General Optimization Guidelines:
@@ -1647,15 +1661,25 @@ def generate_final_report(initial_query: str, context: str, reportstyle: str, le
     word_count = pages * 500
     prompt = (f"""
 // Instructions:
-1. Integrate numbers, quotes, and factual references systematically (We want to incorporate as many relevant numbers, statistics, factual references, quotes from the sources)
-ex: Google CEO ... mentioned that / he report xxx from yyy released in ddd revealed that ... / Mr.xyz said that "..."
 2. Whenever you mention a figure or quote, add an inline reference [x] matching its source from the references.
 3. Again, Specifically name relevant organizations, tools, project names, and people encountered in the crumbs or learnings.
 Note: This is for academic purposes, so thorough citations and referencing are essential.
 4. Focus on reputable sources that will not be disputed (generally social media posts cannot be an opposable sources, but some of them may mention reputable sources)
 Note: put the full reference url (no generic domain address), down to the html page or the pdf
 5. It must follow this writing style {reportstyle}.
-6. All key information giving credit and materiality to the report (numbers, names, people/titles, dates, papers, reports, organisation/institute/NGO/government bodies quotes, products, project names, ... ) should be explicitly mentioned IN the report with ref to the sources from the search results.
 // Sources
 Use the following learnings and merged reference details from a deep research process on:
@@ -1668,6 +1692,7 @@ The report should be very detailed and lengthy — approximately the equivalent
 If you cannot produce a high level of details for specific topic or section, put it in <div class="improvable-chunk">...</div>
 // Requirements
 - It must include inline citations (e.g., [1], [2], etc.) from real sources provided in the search results below
 - Do not add any inline citations reference in the placeholders descriptions below (visual, graph), you can add them in focus though.
 - No more than 7 sentences per div blocks, skip lines and add line breaks when changing topic.
@@ -1824,6 +1849,9 @@ Then close the html code from the broader report
     )
     tokentarget = word_count * 5  # rough multiplier for token target
     report = openai_call(prompt, model="o3-mini", max_tokens_param=tokentarget)
     # If the report is too long, compress it.
     if len(report) > MAX_MESSAGE_LENGTH:
         report = compress_text(report, MAX_MESSAGE_LENGTH)

         padding: 20px;
       }
       h2 {
+        text-align: left;
         margin-bottom: 0;
       }
       #chart {
       font-family: Arial, sans-serif;
     }
     h1 {
+      text-align: left;
       margin: 20px 0;
     }
     #chart {
       background-color: #f5f5f5;
     }
     h1 {
+      text-align: left;
       margin-bottom: 10px;
       color: #333;
     }
 // Important
 - Make the visuals content rich, there's no point having a visual if its content has no real value.
 - It has to convey some relevant insights.
+- Make sure all the items are visible and don't require hovering the mouse to be displayed - this report is meant to be printed.
 - Take a deep breath, think step by step and think it well.
 - Use your judgement to decide between box plots, bubble charts, calendar view, chord diagrams, histograms, ...
 - Your response should start with <html> and end with </html> - no intro before, no comments after.
 - Skip lines and add intermediate titles for key paragraphs
 - You have freedom on the structure, but it has to cover all potential aspects on the topic in between 500 and 1000 words
+// Mentioning sources, organisations and individuals
+- We will perform a post-processing on the output to remove all the identifiable organisations, people and projects
+- For this reasons use this format for any specific name, organisation or project: \{[\{name\}]\}
+ex1: \{[\{Google\}]\} CEO, \{[\{Sundar Pichai\}]\} ...
+ex2: in a report from the \{[\{university of Berkeley\}]\} titled "\{[\{The great acceleration\}]\}"...
+ex3: the CEO of \{[\{Softbank\}]\} , \{[\{Masayoshi Son\}]\}, said that "the best way to..."
+ex4: the project \{[\{Stargate\}]\}, anounced by \{[\{OpenAI\}]\} in collaboration with \{[\{Salesforce\}]\}
+ex5: Mr. \{[\{Michael Parrot\}]\}, Marketing director in \{[\{Panasonic\}]\}, mentioned that ...
+Note: the output will be processed through regex and the identifiers removed, but this way we can keep track of all sources and citations without disclosing them.
+- This should apply to names, people/titles, dates, papers, reports, organisation/institute/NGO/government bodies quotes, products, project names, ...
+Important: you can safely mention any of these provided you respect the above mentioned formatting \{[\{...\}]\}
 // Important
 - Make it real, with anecdotes from the content
 - Ground the content on the inputs shared above (context and knowledge inputs) with facts and numbers
     result = openai_call(prompt, model="o3-mini", max_tokens_param=10000)
     result = result.strip().strip("```").strip()
+    result = re.sub(r'\{\[\{(.*?)\}\]\}', r'\1', result)
     logging.info(f"The code produced for this focus placeholder:\n{placeholder_text}\n\n {result}\n\n")
     return result
     -   Key Facts (at least 5): List the core factual claims using short, declarative sentences or bullet points. Apply lemmatization and standard abbreviations.
     -   Key Figures (at least 5): Extract numerical data (statistics, dates, percentages) and include any necessary context (units, references, explanations) required to interpret these numbers. Present them concisely (list or table format).
     -   Key Arguments (at least 5): Identify main arguments or claims. Summarize supporting evidence and counter-arguments concisely.
+    -   Key Quotes (at least 1 if any): Include significant quotes (with the author's name in parenthesis). Attribute quotes correctly. Paraphrase if needed, indicating that it's a paraphrase. Use symbols (e.g., &, +, ->, =) to conserve tokens.
     -   Structured Summary (10 to 50 sentences): Provide a structured summary that includes anecdotes, people, and locations to ensure the report is relatable.
 Note: General Optimization Guidelines:
     word_count = pages * 500
     prompt = (f"""
 // Instructions:
+1. Integrate numbers from the sources but always mention the source
 2. Whenever you mention a figure or quote, add an inline reference [x] matching its source from the references.
 3. Again, Specifically name relevant organizations, tools, project names, and people encountered in the crumbs or learnings.
 Note: This is for academic purposes, so thorough citations and referencing are essential.
 4. Focus on reputable sources that will not be disputed (generally social media posts cannot be an opposable sources, but some of them may mention reputable sources)
 Note: put the full reference url (no generic domain address), down to the html page or the pdf
 5. It must follow this writing style {reportstyle}.
+// Mentioning sources, organisations and individuals
+- We will perform a post-processing on the output to remove all the identifiable organisations, people and projects
+- For this reasons use this format for any specific name, organisation or project: \{[\{name\}]\}
+ex1: \{[\{Google\}]\} CEO, \{[\{Sundar Pichai\}]\} ...
+ex2: in a report from the \{[\{university of Berkeley\}]\} titled "\{[\{The great acceleration\}]\}"...
+ex3: the CEO of \{[\{Softbank\}]\} , \{[\{Masayoshi Son\}]\}, said that "the best way to..."
+ex4: the project \{[\{Stargate\}]\}, anounced by \{[\{OpenAI\}]\} in collaboration with \{[\{Salesforce\}]\}
+ex5: Mr. \{[\{Michael Parrot\}]\}, Marketing director in \{[\{Panasonic\}]\}, mentioned that ...
+Note: the output will be processed through regex and the identifiers removed, but this way we can keep track of all sources and citations without disclosing them.
+- This should apply to names, people/titles, dates, papers, reports, organisation/institute/NGO/government bodies quotes, products, project names, ...
+Important: you can safely mention any of these provided you respect the above mentioned formatting \{[\{...\}]\}
 // Sources
 Use the following learnings and merged reference details from a deep research process on:
 If you cannot produce a high level of details for specific topic or section, put it in <div class="improvable-chunk">...</div>
 // Requirements
+- All text alignment has to be on the left
 - It must include inline citations (e.g., [1], [2], etc.) from real sources provided in the search results below
 - Do not add any inline citations reference in the placeholders descriptions below (visual, graph), you can add them in focus though.
 - No more than 7 sentences per div blocks, skip lines and add line breaks when changing topic.
     )
     tokentarget = word_count * 5  # rough multiplier for token target
     report = openai_call(prompt, model="o3-mini", max_tokens_param=tokentarget)
+    # Post-processing
+    report = re.sub(r'\{\[\{(.*?)\}\]\}', r'\1', report)
     # If the report is too long, compress it.
     if len(report) > MAX_MESSAGE_LENGTH:
         report = compress_text(report, MAX_MESSAGE_LENGTH)