Saravanakumar R commited on
Commit
e45e84e
·
1 Parent(s): eacbe8f
Files changed (1) hide show
  1. src/agent/agent.py +11 -7
src/agent/agent.py CHANGED
@@ -28,7 +28,6 @@ import os
28
  import time
29
 
30
  import anthropic
31
- from langsmith.wrappers import wrap_anthropic
32
 
33
  from src.agent.memory import SessionMemory
34
  from src.agent.prompts import SYSTEM_PROMPT
@@ -36,7 +35,12 @@ from src.agent.tools import TOOL_DEFINITIONS, execute_tool
36
 
37
  logger = logging.getLogger(__name__)
38
 
39
- _MODEL = "claude-sonnet-4-6"
 
 
 
 
 
40
  _MAX_TOKENS = 4096
41
  # Hard cap on tool-use rounds per turn to prevent runaway loops.
42
  _MAX_TOOL_ROUNDS = 12
@@ -150,20 +154,20 @@ class GUIDEAgent:
150
  # Disable SDK retries so ONLY our _stream_with_retry backoff fires — it
151
  # waits ~60s for the bucket to refill before re-sending.
152
  litellm_url = os.getenv("LITELLM_PROXY_URL")
153
- _tracing = os.getenv("LANGCHAIN_TRACING_V2", "").lower() == "true"
154
  if litellm_url:
155
  from litellm import get_litellm_gateway_api_key
156
- client = anthropic.Anthropic(
157
  base_url=litellm_url,
158
  api_key=get_litellm_gateway_api_key(),
159
  max_retries=0,
160
  )
161
  else:
162
- client = anthropic.Anthropic(
163
- api_key=os.environ["ANTHROPIC_API_KEY"],
 
 
164
  max_retries=0,
165
  )
166
- self._client = wrap_anthropic(client) if _tracing else client
167
 
168
  # ------------------------------------------------------------------
169
  # Public interface
 
28
  import time
29
 
30
  import anthropic
 
31
 
32
  from src.agent.memory import SessionMemory
33
  from src.agent.prompts import SYSTEM_PROMPT
 
35
 
36
  logger = logging.getLogger(__name__)
37
 
38
+ # Model name is env-driven so the same code runs against two backends:
39
+ # • Local via the athena LiteLLM gateway, which only exposes Bedrock model
40
+ # IDs (e.g. bedrock-claude-3-5-haiku-20241022-v1:0) — set GUIDE_MODEL in .env.
41
+ # • HF Spaces / public Anthropic API — leave GUIDE_MODEL unset to use the
42
+ # default claude-sonnet-4-6.
43
+ _MODEL = os.getenv("GUIDE_MODEL", "claude-sonnet-4-6")
44
  _MAX_TOKENS = 4096
45
  # Hard cap on tool-use rounds per turn to prevent runaway loops.
46
  _MAX_TOOL_ROUNDS = 12
 
154
  # Disable SDK retries so ONLY our _stream_with_retry backoff fires — it
155
  # waits ~60s for the bucket to refill before re-sending.
156
  litellm_url = os.getenv("LITELLM_PROXY_URL")
 
157
  if litellm_url:
158
  from litellm import get_litellm_gateway_api_key
159
+ self._client = anthropic.Anthropic(
160
  base_url=litellm_url,
161
  api_key=get_litellm_gateway_api_key(),
162
  max_retries=0,
163
  )
164
  else:
165
+ _key = os.environ.get("ANTHROPIC_API_KEY", "")
166
+ logger.info("GUIDEAgent init key ends=%s len=%d", _key[-3:] if _key else "EMPTY", len(_key))
167
+ self._client = anthropic.Anthropic(
168
+ api_key=_key,
169
  max_retries=0,
170
  )
 
171
 
172
  # ------------------------------------------------------------------
173
  # Public interface