rk147 commited on
Commit
e5e4bf3
·
verified ·
1 Parent(s): 2b21919

Update app/modules/llm.py

Browse files
Files changed (1) hide show
  1. app/modules/llm.py +72 -10
app/modules/llm.py CHANGED
@@ -1,16 +1,21 @@
1
- from groq import Groq
 
2
  import json
 
3
  import streamlit as st
4
- from app.config import GROQ_API_KEY, MODEL_NAME, MAX_TOKENS, TEMPERATURE
 
5
  from app.utils.prompts import (
6
  EXPLAIN_CODE_WITH_LANG, CLASSIFY_ISSUE_WITH_LANG, PROPOSE_FIX_WITH_LANG,
7
- build_chat_prompt
8
  )
9
 
 
10
  @st.cache_resource
11
- def get_client():
12
  return Groq(api_key=GROQ_API_KEY)
13
 
 
14
  def _call(prompt: str, max_tokens: int = MAX_TOKENS) -> str:
15
  client = get_client()
16
  resp = client.chat.completions.create(
@@ -21,6 +26,7 @@ def _call(prompt: str, max_tokens: int = MAX_TOKENS) -> str:
21
  )
22
  return resp.choices[0].message.content.strip()
23
 
 
24
  def _parse_json(text: str) -> dict:
25
  try:
26
  start = text.find("{")
@@ -29,36 +35,92 @@ def _parse_json(text: str) -> dict:
29
  except Exception:
30
  return {"error": "Parse failed", "raw": text[:200]}
31
 
 
32
  def explain_code(code: str, language: str = "python") -> dict:
33
- """Explain code — always uses the language-aware prompt."""
34
  prompt = EXPLAIN_CODE_WITH_LANG.format(language=language, code=code)
35
  return _parse_json(_call(prompt))
36
 
 
37
  def classify_issue(issue: dict, code: str, language: str = "python") -> dict:
38
- """Classify issue — always uses the language-aware prompt."""
39
  prompt = CLASSIFY_ISSUE_WITH_LANG.format(
40
  language=language,
41
  issue=issue.get("message", ""),
42
- code=code
43
  )
44
  return _parse_json(_call(prompt, max_tokens=200))
45
 
 
46
  def propose_fix(issue: dict, code: str, language: str = "python") -> str:
47
- """Propose fix — always uses the language-aware prompt."""
48
  prompt = PROPOSE_FIX_WITH_LANG.format(
49
  language=language,
50
  issue=issue.get("message", ""),
51
- code=code
52
  )
53
  return _call(prompt, max_tokens=512)
54
 
55
- def chat_response(user_query: str, chat_history: list, context: str = "") -> str:
 
 
 
 
 
 
 
 
 
 
 
56
  client = get_client()
57
  system, messages = build_chat_prompt(chat_history, user_query, context)
 
 
 
 
 
 
 
 
 
 
58
  resp = client.chat.completions.create(
59
  model=MODEL_NAME,
60
  messages=[{"role": "system", "content": system}] + messages,
61
  max_tokens=MAX_TOKENS,
62
  temperature=0.3,
63
  )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
64
  return resp.choices[0].message.content.strip()
 
1
+ """LLM interface — text, vision, and chat."""
2
+
3
  import json
4
+ import base64
5
  import streamlit as st
6
+ from groq import Groq
7
+ from app.config import GROQ_API_KEY, MODEL_NAME, VISION_MODEL_NAME, MAX_TOKENS, TEMPERATURE
8
  from app.utils.prompts import (
9
  EXPLAIN_CODE_WITH_LANG, CLASSIFY_ISSUE_WITH_LANG, PROPOSE_FIX_WITH_LANG,
10
+ build_chat_prompt,
11
  )
12
 
13
+
14
  @st.cache_resource
15
+ def get_client() -> Groq:
16
  return Groq(api_key=GROQ_API_KEY)
17
 
18
+
19
  def _call(prompt: str, max_tokens: int = MAX_TOKENS) -> str:
20
  client = get_client()
21
  resp = client.chat.completions.create(
 
26
  )
27
  return resp.choices[0].message.content.strip()
28
 
29
+
30
  def _parse_json(text: str) -> dict:
31
  try:
32
  start = text.find("{")
 
35
  except Exception:
36
  return {"error": "Parse failed", "raw": text[:200]}
37
 
38
+
39
  def explain_code(code: str, language: str = "python") -> dict:
 
40
  prompt = EXPLAIN_CODE_WITH_LANG.format(language=language, code=code)
41
  return _parse_json(_call(prompt))
42
 
43
+
44
  def classify_issue(issue: dict, code: str, language: str = "python") -> dict:
 
45
  prompt = CLASSIFY_ISSUE_WITH_LANG.format(
46
  language=language,
47
  issue=issue.get("message", ""),
48
+ code=code,
49
  )
50
  return _parse_json(_call(prompt, max_tokens=200))
51
 
52
+
53
  def propose_fix(issue: dict, code: str, language: str = "python") -> str:
 
54
  prompt = PROPOSE_FIX_WITH_LANG.format(
55
  language=language,
56
  issue=issue.get("message", ""),
57
+ code=code,
58
  )
59
  return _call(prompt, max_tokens=512)
60
 
61
+
62
+ def chat_response(
63
+ user_query: str,
64
+ chat_history: list,
65
+ context: str = "",
66
+ linked_files: list = None,
67
+ ) -> str:
68
+ """
69
+ Chat response with optional file context.
70
+ Injects a clear system note about available files so model
71
+ never claims it cannot see them.
72
+ """
73
  client = get_client()
74
  system, messages = build_chat_prompt(chat_history, user_query, context)
75
+
76
+ if linked_files:
77
+ file_list = ", ".join(linked_files)
78
+ system += (
79
+ f"\n\n[System: The user has loaded the following files into this session: {file_list}. "
80
+ f"Relevant excerpts from these files are provided above in the context block. "
81
+ f"You have full access to their contents via that context. "
82
+ f"NEVER say you cannot access files or that you are text-based only.]"
83
+ )
84
+
85
  resp = client.chat.completions.create(
86
  model=MODEL_NAME,
87
  messages=[{"role": "system", "content": system}] + messages,
88
  max_tokens=MAX_TOKENS,
89
  temperature=0.3,
90
  )
91
+ return resp.choices[0].message.content.strip()
92
+
93
+
94
+ def vision_response(
95
+ user_query: str,
96
+ image_bytes: bytes,
97
+ media_type: str = "image/png",
98
+ context: str = "",
99
+ ) -> str:
100
+ """Send an image + query to the vision model."""
101
+ client = get_client()
102
+ b64 = base64.b64encode(image_bytes).decode("utf-8")
103
+
104
+ system = "You are an expert software engineer and code analyst. Answer questions about code, architecture diagrams, error screenshots, and related technical images clearly and concisely."
105
+ if context:
106
+ system += f"\n\nAdditional context from linked files:\n```\n{context}\n```"
107
+
108
+ resp = client.chat.completions.create(
109
+ model=VISION_MODEL_NAME,
110
+ messages=[
111
+ {"role": "system", "content": system},
112
+ {
113
+ "role": "user",
114
+ "content": [
115
+ {
116
+ "type": "image_url",
117
+ "image_url": {"url": f"data:{media_type};base64,{b64}"},
118
+ },
119
+ {"type": "text", "text": user_query},
120
+ ],
121
+ },
122
+ ],
123
+ max_tokens=MAX_TOKENS,
124
+ temperature=0.3,
125
+ )
126
  return resp.choices[0].message.content.strip()