CaffeinatedCoding commited on
Commit
2f8ccef
·
verified ·
1 Parent(s): 376c7ca

Upload folder using huggingface_hub

Browse files
Files changed (2) hide show
  1. Dockerfile +0 -0
  2. src/llm.py +14 -3
Dockerfile CHANGED
Binary files a/Dockerfile and b/Dockerfile differ
 
src/llm.py CHANGED
@@ -25,6 +25,16 @@ Rules you must follow:
25
  4. Always cite the Judgment ID when referencing a case
26
  5. If excerpts don't contain enough information, say so explicitly
27
  6. End every response with: "NOTE: This is not legal advice. Consult a qualified advocate."
 
 
 
 
 
 
 
 
 
 
28
  """
29
 
30
  @retry(
@@ -41,7 +51,8 @@ def call_llm(query: str, context: str) -> str:
41
  SUPREME COURT JUDGMENT EXCERPTS:
42
  {context}
43
 
44
- Answer based only on the excerpts above. Cite judgment IDs."""
 
45
 
46
  response = _client.chat.completions.create(
47
  model="llama-3.3-70b-versatile",
@@ -50,7 +61,7 @@ Answer based only on the excerpts above. Cite judgment IDs."""
50
  {"role": "user", "content": user_message}
51
  ],
52
  temperature=0.1,
53
- max_tokens=800
54
  )
55
-
56
  return response.choices[0].message.content
 
25
  4. Always cite the Judgment ID when referencing a case
26
  5. If excerpts don't contain enough information, say so explicitly
27
  6. End every response with: "NOTE: This is not legal advice. Consult a qualified advocate."
28
+
29
+ Formatting rules — follow these exactly:
30
+ - Use numbered lists (1. 2. 3.) when listing multiple points or steps
31
+ - Use bullet points (- item) for sub-points or supporting details
32
+ - Use markdown tables (| Col | Col |) when comparing options side by side
33
+ - Use **bold** for important terms, case names, and section numbers
34
+ - Use headers (## Heading) to separate major sections in long answers
35
+ - Never write everything as one long paragraph
36
+ - Each distinct point must be on its own line
37
+ - Always put a blank line between sections
38
  """
39
 
40
  @retry(
 
51
  SUPREME COURT JUDGMENT EXCERPTS:
52
  {context}
53
 
54
+ Answer based only on the excerpts above. Cite judgment IDs.
55
+ Use proper markdown formatting — numbered lists, bullet points, tables, bold text as appropriate."""
56
 
57
  response = _client.chat.completions.create(
58
  model="llama-3.3-70b-versatile",
 
61
  {"role": "user", "content": user_message}
62
  ],
63
  temperature=0.1,
64
+ max_tokens=1500
65
  )
66
+
67
  return response.choices[0].message.content