File size: 16,449 Bytes
ff5e06a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
import json
import operator
import re
from typing import Annotated, TypedDict

from langchain_core.messages import AIMessage
from langchain_core.tools import render_text_description
from langgraph.graph import END, StateGraph
from langgraph.prebuilt import ToolNode
from modules.tools import (
    get_patient_data_manifest,
    get_patient_fhir_resource,
)

_LLM_INVOKE_ARGS = {"max_tokens": 8000, "temperature": 0.6}


def exclude_thinking_component(text: str) -> str:
  """Removes the thinking block (delimited by <unused94> and <unused95>) from a string."""
  return re.sub(r"<unused94>.*?<unused95>", "", text, flags=re.DOTALL).strip()


def strip_json_decoration(text: str) -> str:
  """Removes JSON markdown fences from the start and end of a string."""
  match = re.search(r"```(?:json)?\s*([\{\[].*[\]\}])\s*```", text, re.DOTALL)
  if match:
    return match.group(1)
  return text.strip()


class AgentState(TypedDict):
  messages: Annotated[list, operator.add]
  patient_fhir_manifest: dict
  tool_output_summary: Annotated[list, operator.add]
  tool_calls_to_execute: Annotated[list, operator.add]
  relevant_resource_types: list
  manifest_tool_call_request: AIMessage
  sdt_idx: int
  edr_idx: int
  resource_type_processed: str
  resource_type_retrieved: str
  summary_generated: bool
  resource_type_to_retrieve: str
  resource_type_to_process: str
  fhir_tool_output: str
  resource_being_summarized: str
  tool_call: dict
  resource_manifest_codes: list


def create_agent(llm, fhir_store_url):
  """Creates and compiles the LangGraph agent."""

  manifest_tool_node = ToolNode([get_patient_data_manifest])
  data_retrieval_tool_node = ToolNode([get_patient_fhir_resource])

  def generate_manifest_tool_call_node(state):
    """The first step: uses the LLM to find the patient_id from the initial question

    and generates a tool call for get_patient_data_manifest.
    """
    last_message = state["messages"][-1]
    extraction_prompt = (
        f"USER QUESTION: {last_message.content}\\n\\nYou are an API request"
        " generator. Your task is to identify the patient ID from the user's"
        " question and output a JSON object to call the"
        " `get_patient_data_manifest` tool.\\n\\nYour available tool"
        f" is:\\n{render_text_description([get_patient_data_manifest])}\\n\\nGenerate"
        " the correct JSON to call the tool. Respond with only a single, raw"
        ' JSON object.\\n\\nEXAMPLE:\\n{\\n  "name":'
        ' "get_patient_data_manifest",\\n  "args": {\\n    "patient_id":'
        ' "some-patient-id-from-the-question"\\n  }\\n}\\n'
    )
    print(
        "--- generate_manifest_tool_call_node PROMPT"
        f" ---\n{extraction_prompt}\n-----------------------------"
    )
    response_str = llm.invoke(extraction_prompt, **_LLM_INVOKE_ARGS)
    print(
        "--- generate_manifest_tool_call_node RESPONSE"
        f" ---\n{response_str}\n------------------------------"
    )
    try:
      cleaned_response = strip_json_decoration(response_str)
      tool_call_json = json.loads(cleaned_response)
      tool_call_json["args"]["fhir_store_url"] = fhir_store_url
      tool_call_msg = AIMessage(
          content="",
          tool_calls=[{**tool_call_json, "id": "manifest_call"}],
      )
      return {
          "manifest_tool_call_request": tool_call_msg,
          "tool_call": tool_call_json,
      }
    except Exception as e:
      print(f"Error generating manifest tool call: {e}")
      raise e

  def execute_manifest_tool_call_node(state):
    """Executes the get_patient_data_manifest tool call and puts the result in state."""
    try:
      tool_call_msg = state["manifest_tool_call_request"]
      tool_output_message = manifest_tool_node.invoke([tool_call_msg])[0]
      manifest_dict = json.loads(tool_output_message.content)
      print(f"Manifest dict: {manifest_dict}")
      return {"patient_fhir_manifest": manifest_dict}
    except Exception as e:
      print(f"Error calling manifest tool: {e}")
      raise e

  def identify_relevant_resource_types(state):
    """Uses the manifest and user question to identify relevant FHIR resource types."""
    print("Identifying Relevant Resource Types")
    manifest = state.get("patient_fhir_manifest", {})
    user_question = state["messages"][1].content
    manifest_content = ""
    for resource_type, codes in manifest.items():
      manifest_content += f"**{resource_type}**: "
      if codes:
        manifest_content += f"Available codes include: {', '.join(codes)}\\n"
      else:
        manifest_content += "Present (no specific codes found)\\n"
    prompt = (
        "SYSTEM INSTRUCTION: think silently if needed.\\nUSER QUESTION:"
        f" {user_question}\\n\\nPATIENT DATA"
        f" MANIFEST:\\n{manifest_content}\\n\\nYou are a medical assistant"
        " analyzing a patient's FHIR data manifest to answer a user"
        " question.\\nBased on the user question, identify the specific FHIR"
        " resource types from the manifest that are most likely to contain the"
        " information needed to answer the question.\\nOutput a JSON list of"
        " the relevant resource types. Do not include any other text or"
        ' formatting.\\nExample:\n["Condition", "Observation",'
        ' "MedicationRequest"]\n'
    )
    print(
        "--- identify_relevant_resource_types PROMPT"
        f" ---\n{prompt}\n------------------------------------------"
    )
    response_str = llm.invoke(prompt, **_LLM_INVOKE_ARGS)
    print(
        "--- identify_relevant_resource_types RESPONSE"
        f" ---\n{response_str}\n-------------------------------------------"
    )
    try:
      relevant_resource_types = json.loads(strip_json_decoration(response_str))
    except json.JSONDecodeError:
      print(
          "Could not decode JSON response for relevant resource types:"
          f" {response_str}"
      )
      relevant_resource_types = []
    print(
        "Relevant Resource Types Identified:"
        f" {', '.join(relevant_resource_types)}"
    )
    return {
        "relevant_resource_types": relevant_resource_types,
        "sdt_idx": 0,
        "tool_calls_to_execute": [],
    }

  def announce_sdt_node(state):
    sdt_idx = state["sdt_idx"]
    relevant_resource_types = state.get("relevant_resource_types", [])
    resource_type = relevant_resource_types[sdt_idx]
    manifest = state.get("patient_fhir_manifest", {})
    resource_manifest = manifest.get(resource_type, [])
    print(f"Announcing data selection for {resource_type}")
    return {
        "resource_type_to_process": resource_type,
        "resource_manifest_codes": resource_manifest,
    }

  def select_data_to_retrieve(state):
    """Uses the manifest and relevant resource types to determine which FHIR resources to retrieve."""
    sdt_idx = state["sdt_idx"]
    manifest = state.get("patient_fhir_manifest", {})
    relevant_resource_types = state.get("relevant_resource_types", [])
    tools_string = render_text_description([get_patient_fhir_resource])

    resource_type = relevant_resource_types[sdt_idx]
    print(f"Data Selection for {resource_type}")

    if resource_type not in manifest:
      print(f"No data found for {resource_type} in the manifest.")
      return {"sdt_idx": sdt_idx + 1, "resource_type_processed": resource_type}

    manifest_content = f"**{resource_type}**: "
    if len(manifest.get(resource_type, [])) > 0:
      manifest_content += (
          f"Available codes include: {', '.join(manifest[resource_type])}\\n"
      )
    else:
      manifest_content += "Present (no specific codes found)\\n"
    prompt = (
        "SYSTEM INSTRUCTION: think silently if needed.\\n"
        + "FOR CONTEXT ONLY, USER QUESTION:"
        f" {state['messages'][1].content}\\n\\n"
        + f"PATIENT DATA MANIFEST: {manifest_content}\\n\\n"
        + "You are a specialized API request generator. Your SOLE task is to"
        " output a JSON of a tool call to gather the necessary information"
        " to answer the user's question. Respond with ONLY a JSON, no"
        " explanations or prose.\\n"
        + f"Your available tool is:\\n{tools_string}\\n\\n"
        + f"**At this stage you can only call {resource_type}.**\\n"
        + "EXAMPLE:\\n"
        + '{\\"name\\": \\"get_patient_fhir_resource\\", \\"args\\":'
        ' {\\"patient_id\\": \\"some-patient-id\\",'
        ' \\"fhir_resource\\": \\"'
        + resource_type
        + '\\", \\"filter_code\\": \\"csv-codes-from-manifest\\"}}'
    )
    print(
        f"--- select_data_to_retrieve PROMPT ({resource_type})"
        f" ---\n{prompt}\n------------------------------------------"
    )
    response_str = llm.invoke(prompt, **_LLM_INVOKE_ARGS)
    print(
        f"--- select_data_to_retrieve RESPONSE ({resource_type})"
        f" ---\n{response_str}\n-------------------------------------------"
    )
    try:
      tool_call = json.loads(strip_json_decoration(response_str))
      tool_call["args"]["fhir_store_url"] = fhir_store_url
      return {
          "tool_calls_to_execute": [{**tool_call, "id": resource_type}],
          "sdt_idx": sdt_idx + 1,
          "resource_type_processed": resource_type,
      }
    except json.JSONDecodeError:
      print(
          f"Could not decode JSON response for {resource_type}: {response_str}"
      )
      # If we fail to decode, we just skip this resource type.
      return {"sdt_idx": sdt_idx + 1, "resource_type_processed": resource_type}

  def sdt_conditional_edge(state):
    if state["sdt_idx"] < len(state["relevant_resource_types"]):
      return "announce_sdt"
    return "init_edr_idx"

  def init_edr_idx_node(state):
    return {"edr_idx": 0}

  def init_edr_conditional_edge(state):
    if state["tool_calls_to_execute"]:
      return "announce_retrieval"
    return "final_answer"

  def announce_retrieval_node(state):
    edr_idx = state["edr_idx"]
    tool_calls = state.get("tool_calls_to_execute", [])
    tool_call = tool_calls[edr_idx]
    resource_type = tool_call.get("id", "unknown_resource")
    print(f"Announcing retrieval of {resource_type}")
    return {"resource_type_to_retrieve": resource_type}

  def execute_data_retrieval(state):
    """Executes the planned tool calls and summarizes the output."""
    edr_idx = state["edr_idx"]
    tool_calls = state.get("tool_calls_to_execute", [])
    tool_call = tool_calls[edr_idx]
    resource_type = tool_call.get("id", "unknown_resource")
    print(f"Fetching FHIR data for {resource_type}")
    tool_output_list = data_retrieval_tool_node.invoke(
        [AIMessage(content="", tool_calls=[tool_call])]
    )
    if not tool_output_list:
      print(f"No tool output received for {resource_type}")
      return {
          "resource_type_retrieved": resource_type,
          "summary_generated": False,
          "fhir_tool_output": "",
      }

    tool_output = tool_output_list[0].content
    return {
        "resource_type_retrieved": resource_type,
        "summary_generated": True,
        "fhir_tool_output": tool_output,
    }

  def announce_summarization_node(state):
    resource_type = state["resource_type_retrieved"]
    print(f"Announcing summarization of {resource_type}")
    return {"resource_being_summarized": resource_type}

  def summarize_node(state: AgentState) -> dict:
    if not state["summary_generated"]:
      return {"edr_idx": state["edr_idx"] + 1}

    resource_type = state["resource_type_retrieved"]
    tool_output = state["fhir_tool_output"]
    concise_facts_prompt = (
        "SYSTEM INSTRUCTION: think silently if needed.\\nFOR CONTEXT ONLY,"
        f" USER QUESTION: {state['messages'][1].content}\\n\\nTOOL"
        f" OUTPUT:\\n{tool_output}\\n\\nYou are a fact summarizing agent."
        " Your output will be used to answer the USER QUESTION.\\nCollect"
        " from the 'TOOL OUTPUT' facts ONLY if it is relevant to answer the"
        " USER QUESTION.\\nWrite a very concise English summary, only facts"
        " relevant to the user question. DO NOT OUTPUT JSON.\\nYou are not"
        " authorized to answer the user question. Do not provide any output"
        " beyond concise facts. Filter out any facts which are not helpful"
        " for the user question. Include date or date ranges. Only for the"
        " most critical facts, include FHIR record references [record"
        " type/record id]. For repeating multiple times provide summarize"
        " and provide only a single reference and date range."
    )
    print(
        f"--- summarize_node PROMPT ({resource_type})"
        f" ---\n{concise_facts_prompt}\n------------------------------------------"
    )
    current_summary = llm.invoke(concise_facts_prompt, **_LLM_INVOKE_ARGS)
    print(
        f"--- summarize_node RESPONSE ({resource_type})"
        f" ---\n{current_summary}\n-------------------------------------------"
    )
    return {
        "tool_output_summary": [exclude_thinking_component(current_summary)],
        "edr_idx": state["edr_idx"] + 1,
        "resource_type_retrieved": resource_type,
    }

  def should_summarize_edge(state):
    if state["summary_generated"]:
      return "announce_summarization"
    return "summarize_node"

  def edr_conditional_edge(state):
    if state["edr_idx"] < len(state["tool_calls_to_execute"]):
      return "announce_retrieval"
    return "final_answer"

  def get_final_answer(state):
    """If we have enough data, this node generates the final answer."""
    summary = "\\n\\n".join(state["tool_output_summary"])
    prompt = (
        "Synthesize all information from the 'SUMMARIZED INFORMATION' to"
        " provide a comprehensive final answer. Preserve relevant FHIR"
        " references.\\n\\nUSER QUESTION:"
        f" {state['messages'][1].content}\\n\\nSUMMARIZED INFORMATION:"
        f" {summary}\\n\\nFinal Answer using markdown:"
    )
    print(
        "--- get_final_answer PROMPT"
        f" ---\n{prompt}\n------------------------------------------"
    )
    response = llm.invoke(prompt, **_LLM_INVOKE_ARGS)
    print(
        "--- get_final_answer RESPONSE"
        f" ---\n{response}\n-------------------------------------------"
    )
    return {"messages": [AIMessage(content=response)]}

  workflow = StateGraph(AgentState)
  workflow.add_node(
      "generate_manifest_tool_call", generate_manifest_tool_call_node
  )
  workflow.add_node(
      "execute_manifest_tool_call", execute_manifest_tool_call_node
  )
  workflow.add_node(
      "identify_relevant_resource_types", identify_relevant_resource_types
  )
  workflow.add_node("announce_sdt", announce_sdt_node)
  workflow.add_node("select_data_to_retrieve", select_data_to_retrieve)
  workflow.add_node("init_edr_idx", init_edr_idx_node)
  workflow.add_node("announce_retrieval", announce_retrieval_node)
  workflow.add_node("execute_data_retrieval", execute_data_retrieval)
  workflow.add_node("announce_summarization", announce_summarization_node)
  workflow.add_node("summarize_node", summarize_node)
  workflow.add_node("final_answer", get_final_answer)
  workflow.set_entry_point("generate_manifest_tool_call")
  workflow.add_edge("generate_manifest_tool_call", "execute_manifest_tool_call")
  workflow.add_edge(
      "execute_manifest_tool_call", "identify_relevant_resource_types"
  )
  workflow.add_edge(
      "identify_relevant_resource_types", "announce_sdt"
  )
  workflow.add_edge("announce_sdt", "select_data_to_retrieve")
  workflow.add_conditional_edges(
      "select_data_to_retrieve",
      sdt_conditional_edge,
      {
          "announce_sdt": "announce_sdt",
          "init_edr_idx": "init_edr_idx",
      },
  )
  workflow.add_conditional_edges(
      "init_edr_idx",
      init_edr_conditional_edge,
      {
          "announce_retrieval": "announce_retrieval",
          "final_answer": "final_answer",
      },
  )
  workflow.add_edge("announce_retrieval", "execute_data_retrieval")
  workflow.add_conditional_edges(
      "execute_data_retrieval",
      should_summarize_edge,
      {
          "announce_summarization": "announce_summarization",
          "summarize_node": "summarize_node",
      },
  )
  workflow.add_edge("announce_summarization", "summarize_node")
  workflow.add_conditional_edges(
      "summarize_node",
      edr_conditional_edge,
      {
          "announce_retrieval": "announce_retrieval",
          "final_answer": "final_answer",
      },
  )
  workflow.add_edge("final_answer", END)
  return workflow.compile()