middleware.py 74 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939
  1. import time
  2. import logging
  3. import sys
  4. import os
  5. import base64
  6. import asyncio
  7. from aiocache import cached
  8. from typing import Any, Optional
  9. import random
  10. import json
  11. import html
  12. import inspect
  13. import re
  14. import ast
  15. from uuid import uuid4
  16. from concurrent.futures import ThreadPoolExecutor
  17. from fastapi import Request
  18. from fastapi import BackgroundTasks
  19. from starlette.responses import Response, StreamingResponse
  20. from open_webui.models.chats import Chats
  21. from open_webui.models.users import Users
  22. from open_webui.socket.main import (
  23. get_event_call,
  24. get_event_emitter,
  25. get_active_status_by_user_id,
  26. )
  27. from open_webui.routers.tasks import (
  28. generate_queries,
  29. generate_title,
  30. generate_image_prompt,
  31. generate_chat_tags,
  32. )
  33. from open_webui.routers.retrieval import process_web_search, SearchForm
  34. from open_webui.routers.images import image_generations, GenerateImageForm
  35. from open_webui.routers.pipelines import (
  36. process_pipeline_inlet_filter,
  37. process_pipeline_outlet_filter,
  38. )
  39. from open_webui.utils.webhook import post_webhook
  40. from open_webui.models.users import UserModel
  41. from open_webui.models.functions import Functions
  42. from open_webui.models.models import Models
  43. from open_webui.retrieval.utils import get_sources_from_files
  44. from open_webui.utils.chat import generate_chat_completion
  45. from open_webui.utils.task import (
  46. get_task_model_id,
  47. rag_template,
  48. tools_function_calling_generation_template,
  49. )
  50. from open_webui.utils.misc import (
  51. deep_update,
  52. get_message_list,
  53. add_or_update_system_message,
  54. add_or_update_user_message,
  55. get_last_user_message,
  56. get_last_assistant_message,
  57. prepend_to_first_user_message_content,
  58. )
  59. from open_webui.utils.tools import get_tools
  60. from open_webui.utils.plugin import load_function_module_by_id
  61. from open_webui.utils.filter import (
  62. get_sorted_filter_ids,
  63. process_filter_functions,
  64. )
  65. from open_webui.utils.code_interpreter import execute_code_jupyter
  66. from open_webui.tasks import create_task
  67. from open_webui.config import (
  68. CACHE_DIR,
  69. DEFAULT_TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE,
  70. DEFAULT_CODE_INTERPRETER_PROMPT,
  71. )
  72. from open_webui.env import (
  73. SRC_LOG_LEVELS,
  74. GLOBAL_LOG_LEVEL,
  75. BYPASS_MODEL_ACCESS_CONTROL,
  76. ENABLE_REALTIME_CHAT_SAVE,
  77. )
  78. from open_webui.constants import TASKS
  79. logging.basicConfig(stream=sys.stdout, level=GLOBAL_LOG_LEVEL)
  80. log = logging.getLogger(__name__)
  81. log.setLevel(SRC_LOG_LEVELS["MAIN"])
  82. async def chat_completion_tools_handler(
  83. request: Request, body: dict, user: UserModel, models, tools
  84. ) -> tuple[dict, dict]:
  85. async def get_content_from_response(response) -> Optional[str]:
  86. content = None
  87. if hasattr(response, "body_iterator"):
  88. async for chunk in response.body_iterator:
  89. data = json.loads(chunk.decode("utf-8"))
  90. content = data["choices"][0]["message"]["content"]
  91. # Cleanup any remaining background tasks if necessary
  92. if response.background is not None:
  93. await response.background()
  94. else:
  95. content = response["choices"][0]["message"]["content"]
  96. return content
  97. def get_tools_function_calling_payload(messages, task_model_id, content):
  98. user_message = get_last_user_message(messages)
  99. history = "\n".join(
  100. f"{message['role'].upper()}: \"\"\"{message['content']}\"\"\""
  101. for message in messages[::-1][:4]
  102. )
  103. prompt = f"History:\n{history}\nQuery: {user_message}"
  104. return {
  105. "model": task_model_id,
  106. "messages": [
  107. {"role": "system", "content": content},
  108. {"role": "user", "content": f"Query: {prompt}"},
  109. ],
  110. "stream": False,
  111. "metadata": {"task": str(TASKS.FUNCTION_CALLING)},
  112. }
  113. task_model_id = get_task_model_id(
  114. body["model"],
  115. request.app.state.config.TASK_MODEL,
  116. request.app.state.config.TASK_MODEL_EXTERNAL,
  117. models,
  118. )
  119. skip_files = False
  120. sources = []
  121. specs = [tool["spec"] for tool in tools.values()]
  122. tools_specs = json.dumps(specs)
  123. if request.app.state.config.TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE != "":
  124. template = request.app.state.config.TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE
  125. else:
  126. template = DEFAULT_TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE
  127. tools_function_calling_prompt = tools_function_calling_generation_template(
  128. template, tools_specs
  129. )
  130. log.info(f"{tools_function_calling_prompt=}")
  131. payload = get_tools_function_calling_payload(
  132. body["messages"], task_model_id, tools_function_calling_prompt
  133. )
  134. try:
  135. response = await generate_chat_completion(request, form_data=payload, user=user)
  136. log.debug(f"{response=}")
  137. content = await get_content_from_response(response)
  138. log.debug(f"{content=}")
  139. if not content:
  140. return body, {}
  141. try:
  142. content = content[content.find("{") : content.rfind("}") + 1]
  143. if not content:
  144. raise Exception("No JSON object found in the response")
  145. result = json.loads(content)
  146. async def tool_call_handler(tool_call):
  147. nonlocal skip_files
  148. log.debug(f"{tool_call=}")
  149. tool_function_name = tool_call.get("name", None)
  150. if tool_function_name not in tools:
  151. return body, {}
  152. tool_function_params = tool_call.get("parameters", {})
  153. try:
  154. required_params = (
  155. tools[tool_function_name]
  156. .get("spec", {})
  157. .get("parameters", {})
  158. .get("required", [])
  159. )
  160. tool_function = tools[tool_function_name]["callable"]
  161. tool_function_params = {
  162. k: v
  163. for k, v in tool_function_params.items()
  164. if k in required_params
  165. }
  166. tool_output = await tool_function(**tool_function_params)
  167. except Exception as e:
  168. tool_output = str(e)
  169. if isinstance(tool_output, str):
  170. if tools[tool_function_name]["citation"]:
  171. sources.append(
  172. {
  173. "source": {
  174. "name": f"TOOL:{tools[tool_function_name]['toolkit_id']}/{tool_function_name}"
  175. },
  176. "document": [tool_output],
  177. "metadata": [
  178. {
  179. "source": f"TOOL:{tools[tool_function_name]['toolkit_id']}/{tool_function_name}"
  180. }
  181. ],
  182. }
  183. )
  184. else:
  185. sources.append(
  186. {
  187. "source": {},
  188. "document": [tool_output],
  189. "metadata": [
  190. {
  191. "source": f"TOOL:{tools[tool_function_name]['toolkit_id']}/{tool_function_name}"
  192. }
  193. ],
  194. }
  195. )
  196. if tools[tool_function_name]["file_handler"]:
  197. skip_files = True
  198. # check if "tool_calls" in result
  199. if result.get("tool_calls"):
  200. for tool_call in result.get("tool_calls"):
  201. await tool_call_handler(tool_call)
  202. else:
  203. await tool_call_handler(result)
  204. except Exception as e:
  205. log.exception(f"Error: {e}")
  206. content = None
  207. except Exception as e:
  208. log.exception(f"Error: {e}")
  209. content = None
  210. log.debug(f"tool_contexts: {sources}")
  211. if skip_files and "files" in body.get("metadata", {}):
  212. del body["metadata"]["files"]
  213. return body, {"sources": sources}
  214. async def chat_web_search_handler(
  215. request: Request, form_data: dict, extra_params: dict, user
  216. ):
  217. event_emitter = extra_params["__event_emitter__"]
  218. await event_emitter(
  219. {
  220. "type": "status",
  221. "data": {
  222. "action": "web_search",
  223. "description": "Generating search query",
  224. "done": False,
  225. },
  226. }
  227. )
  228. messages = form_data["messages"]
  229. user_message = get_last_user_message(messages)
  230. queries = []
  231. try:
  232. res = await generate_queries(
  233. request,
  234. {
  235. "model": form_data["model"],
  236. "messages": messages,
  237. "prompt": user_message,
  238. "type": "web_search",
  239. },
  240. user,
  241. )
  242. response = res["choices"][0]["message"]["content"]
  243. try:
  244. bracket_start = response.find("{")
  245. bracket_end = response.rfind("}") + 1
  246. if bracket_start == -1 or bracket_end == -1:
  247. raise Exception("No JSON object found in the response")
  248. response = response[bracket_start:bracket_end]
  249. queries = json.loads(response)
  250. queries = queries.get("queries", [])
  251. except Exception as e:
  252. queries = [response]
  253. except Exception as e:
  254. log.exception(e)
  255. queries = [user_message]
  256. if len(queries) == 0:
  257. await event_emitter(
  258. {
  259. "type": "status",
  260. "data": {
  261. "action": "web_search",
  262. "description": "No search query generated",
  263. "done": True,
  264. },
  265. }
  266. )
  267. return form_data
  268. searchQuery = queries[0]
  269. await event_emitter(
  270. {
  271. "type": "status",
  272. "data": {
  273. "action": "web_search",
  274. "description": 'Searching "{{searchQuery}}"',
  275. "query": searchQuery,
  276. "done": False,
  277. },
  278. }
  279. )
  280. try:
  281. results = await process_web_search(
  282. request,
  283. SearchForm(
  284. **{
  285. "query": searchQuery,
  286. }
  287. ),
  288. user,
  289. )
  290. if results:
  291. await event_emitter(
  292. {
  293. "type": "status",
  294. "data": {
  295. "action": "web_search",
  296. "description": "Searched {{count}} sites",
  297. "query": searchQuery,
  298. "urls": results["filenames"],
  299. "done": True,
  300. },
  301. }
  302. )
  303. files = form_data.get("files", [])
  304. files.append(
  305. {
  306. "collection_name": results["collection_name"],
  307. "name": searchQuery,
  308. "type": "web_search_results",
  309. "urls": results["filenames"],
  310. }
  311. )
  312. form_data["files"] = files
  313. else:
  314. await event_emitter(
  315. {
  316. "type": "status",
  317. "data": {
  318. "action": "web_search",
  319. "description": "No search results found",
  320. "query": searchQuery,
  321. "done": True,
  322. "error": True,
  323. },
  324. }
  325. )
  326. except Exception as e:
  327. log.exception(e)
  328. await event_emitter(
  329. {
  330. "type": "status",
  331. "data": {
  332. "action": "web_search",
  333. "description": 'Error searching "{{searchQuery}}"',
  334. "query": searchQuery,
  335. "done": True,
  336. "error": True,
  337. },
  338. }
  339. )
  340. return form_data
  341. async def chat_image_generation_handler(
  342. request: Request, form_data: dict, extra_params: dict, user
  343. ):
  344. __event_emitter__ = extra_params["__event_emitter__"]
  345. await __event_emitter__(
  346. {
  347. "type": "status",
  348. "data": {"description": "Generating an image", "done": False},
  349. }
  350. )
  351. messages = form_data["messages"]
  352. user_message = get_last_user_message(messages)
  353. prompt = user_message
  354. negative_prompt = ""
  355. if request.app.state.config.ENABLE_IMAGE_PROMPT_GENERATION:
  356. try:
  357. res = await generate_image_prompt(
  358. request,
  359. {
  360. "model": form_data["model"],
  361. "messages": messages,
  362. },
  363. user,
  364. )
  365. response = res["choices"][0]["message"]["content"]
  366. try:
  367. bracket_start = response.find("{")
  368. bracket_end = response.rfind("}") + 1
  369. if bracket_start == -1 or bracket_end == -1:
  370. raise Exception("No JSON object found in the response")
  371. response = response[bracket_start:bracket_end]
  372. response = json.loads(response)
  373. prompt = response.get("prompt", [])
  374. except Exception as e:
  375. prompt = user_message
  376. except Exception as e:
  377. log.exception(e)
  378. prompt = user_message
  379. system_message_content = ""
  380. try:
  381. images = await image_generations(
  382. request=request,
  383. form_data=GenerateImageForm(**{"prompt": prompt}),
  384. user=user,
  385. )
  386. await __event_emitter__(
  387. {
  388. "type": "status",
  389. "data": {"description": "Generated an image", "done": True},
  390. }
  391. )
  392. for image in images:
  393. await __event_emitter__(
  394. {
  395. "type": "message",
  396. "data": {"content": f"![Generated Image]({image['url']})\n"},
  397. }
  398. )
  399. system_message_content = "<context>User is shown the generated image, tell the user that the image has been generated</context>"
  400. except Exception as e:
  401. log.exception(e)
  402. await __event_emitter__(
  403. {
  404. "type": "status",
  405. "data": {
  406. "description": f"An error occurred while generating an image",
  407. "done": True,
  408. },
  409. }
  410. )
  411. system_message_content = "<context>Unable to generate an image, tell the user that an error occurred</context>"
  412. if system_message_content:
  413. form_data["messages"] = add_or_update_system_message(
  414. system_message_content, form_data["messages"]
  415. )
  416. return form_data
  417. async def chat_completion_files_handler(
  418. request: Request, body: dict, user: UserModel
  419. ) -> tuple[dict, dict[str, list]]:
  420. sources = []
  421. if files := body.get("metadata", {}).get("files", None):
  422. try:
  423. queries_response = await generate_queries(
  424. request,
  425. {
  426. "model": body["model"],
  427. "messages": body["messages"],
  428. "type": "retrieval",
  429. },
  430. user,
  431. )
  432. queries_response = queries_response["choices"][0]["message"]["content"]
  433. try:
  434. bracket_start = queries_response.find("{")
  435. bracket_end = queries_response.rfind("}") + 1
  436. if bracket_start == -1 or bracket_end == -1:
  437. raise Exception("No JSON object found in the response")
  438. queries_response = queries_response[bracket_start:bracket_end]
  439. queries_response = json.loads(queries_response)
  440. except Exception as e:
  441. queries_response = {"queries": [queries_response]}
  442. queries = queries_response.get("queries", [])
  443. except Exception as e:
  444. queries = []
  445. if len(queries) == 0:
  446. queries = [get_last_user_message(body["messages"])]
  447. try:
  448. # Offload get_sources_from_files to a separate thread
  449. loop = asyncio.get_running_loop()
  450. with ThreadPoolExecutor() as executor:
  451. sources = await loop.run_in_executor(
  452. executor,
  453. lambda: get_sources_from_files(
  454. files=files,
  455. queries=queries,
  456. embedding_function=lambda query: request.app.state.EMBEDDING_FUNCTION(
  457. query, user=user
  458. ),
  459. k=request.app.state.config.TOP_K,
  460. reranking_function=request.app.state.rf,
  461. r=request.app.state.config.RELEVANCE_THRESHOLD,
  462. hybrid_search=request.app.state.config.ENABLE_RAG_HYBRID_SEARCH,
  463. ),
  464. )
  465. except Exception as e:
  466. log.exception(e)
  467. log.debug(f"rag_contexts:sources: {sources}")
  468. return body, {"sources": sources}
  469. def apply_params_to_form_data(form_data, model):
  470. params = form_data.pop("params", {})
  471. if model.get("ollama"):
  472. form_data["options"] = params
  473. if "format" in params:
  474. form_data["format"] = params["format"]
  475. if "keep_alive" in params:
  476. form_data["keep_alive"] = params["keep_alive"]
  477. else:
  478. if "seed" in params:
  479. form_data["seed"] = params["seed"]
  480. if "stop" in params:
  481. form_data["stop"] = params["stop"]
  482. if "temperature" in params:
  483. form_data["temperature"] = params["temperature"]
  484. if "max_tokens" in params:
  485. form_data["max_tokens"] = params["max_tokens"]
  486. if "top_p" in params:
  487. form_data["top_p"] = params["top_p"]
  488. if "frequency_penalty" in params:
  489. form_data["frequency_penalty"] = params["frequency_penalty"]
  490. if "reasoning_effort" in params:
  491. form_data["reasoning_effort"] = params["reasoning_effort"]
  492. if "logit_bias" in params:
  493. try:
  494. form_data["logit_bias"] = json.loads(params["logit_bias"])
  495. except json.JSONDecodeError:
  496. print("Invalid JSON format for logit_bias")
  497. return form_data
  498. async def process_chat_payload(request, form_data, metadata, user, model):
  499. form_data = apply_params_to_form_data(form_data, model)
  500. log.debug(f"form_data: {form_data}")
  501. event_emitter = get_event_emitter(metadata)
  502. event_call = get_event_call(metadata)
  503. extra_params = {
  504. "__event_emitter__": event_emitter,
  505. "__event_call__": event_call,
  506. "__user__": {
  507. "id": user.id,
  508. "email": user.email,
  509. "name": user.name,
  510. "role": user.role,
  511. },
  512. "__metadata__": metadata,
  513. "__request__": request,
  514. "__model__": model,
  515. }
  516. # Initialize events to store additional event to be sent to the client
  517. # Initialize contexts and citation
  518. if getattr(request.state, "direct", False) and hasattr(request.state, "model"):
  519. models = {
  520. request.state.model["id"]: request.state.model,
  521. }
  522. else:
  523. models = request.app.state.MODELS
  524. task_model_id = get_task_model_id(
  525. form_data["model"],
  526. request.app.state.config.TASK_MODEL,
  527. request.app.state.config.TASK_MODEL_EXTERNAL,
  528. models,
  529. )
  530. events = []
  531. sources = []
  532. user_message = get_last_user_message(form_data["messages"])
  533. model_knowledge = model.get("info", {}).get("meta", {}).get("knowledge", False)
  534. if model_knowledge:
  535. await event_emitter(
  536. {
  537. "type": "status",
  538. "data": {
  539. "action": "knowledge_search",
  540. "query": user_message,
  541. "done": False,
  542. },
  543. }
  544. )
  545. knowledge_files = []
  546. for item in model_knowledge:
  547. if item.get("collection_name"):
  548. knowledge_files.append(
  549. {
  550. "id": item.get("collection_name"),
  551. "name": item.get("name"),
  552. "legacy": True,
  553. }
  554. )
  555. elif item.get("collection_names"):
  556. knowledge_files.append(
  557. {
  558. "name": item.get("name"),
  559. "type": "collection",
  560. "collection_names": item.get("collection_names"),
  561. "legacy": True,
  562. }
  563. )
  564. else:
  565. knowledge_files.append(item)
  566. files = form_data.get("files", [])
  567. files.extend(knowledge_files)
  568. form_data["files"] = files
  569. variables = form_data.pop("variables", None)
  570. # Process the form_data through the pipeline
  571. try:
  572. form_data = await process_pipeline_inlet_filter(
  573. request, form_data, user, models
  574. )
  575. except Exception as e:
  576. raise e
  577. try:
  578. form_data, flags = await process_filter_functions(
  579. request=request,
  580. filter_ids=get_sorted_filter_ids(model),
  581. filter_type="inlet",
  582. form_data=form_data,
  583. extra_params=extra_params,
  584. )
  585. except Exception as e:
  586. raise Exception(f"Error: {e}")
  587. features = form_data.pop("features", None)
  588. if features:
  589. if "web_search" in features and features["web_search"]:
  590. form_data = await chat_web_search_handler(
  591. request, form_data, extra_params, user
  592. )
  593. if "image_generation" in features and features["image_generation"]:
  594. form_data = await chat_image_generation_handler(
  595. request, form_data, extra_params, user
  596. )
  597. if "code_interpreter" in features and features["code_interpreter"]:
  598. form_data["messages"] = add_or_update_user_message(
  599. (
  600. request.app.state.config.CODE_INTERPRETER_PROMPT_TEMPLATE
  601. if request.app.state.config.CODE_INTERPRETER_PROMPT_TEMPLATE != ""
  602. else DEFAULT_CODE_INTERPRETER_PROMPT
  603. ),
  604. form_data["messages"],
  605. )
  606. tool_ids = form_data.pop("tool_ids", None)
  607. files = form_data.pop("files", None)
  608. # Remove files duplicates
  609. if files:
  610. files = list({json.dumps(f, sort_keys=True): f for f in files}.values())
  611. metadata = {
  612. **metadata,
  613. "tool_ids": tool_ids,
  614. "files": files,
  615. }
  616. form_data["metadata"] = metadata
  617. tool_ids = metadata.get("tool_ids", None)
  618. log.debug(f"{tool_ids=}")
  619. if tool_ids:
  620. # If tool_ids field is present, then get the tools
  621. tools = get_tools(
  622. request,
  623. tool_ids,
  624. user,
  625. {
  626. **extra_params,
  627. "__model__": models[task_model_id],
  628. "__messages__": form_data["messages"],
  629. "__files__": metadata.get("files", []),
  630. },
  631. )
  632. log.info(f"{tools=}")
  633. if metadata.get("function_calling") == "native":
  634. # If the function calling is native, then call the tools function calling handler
  635. metadata["tools"] = tools
  636. form_data["tools"] = [
  637. {"type": "function", "function": tool.get("spec", {})}
  638. for tool in tools.values()
  639. ]
  640. else:
  641. # If the function calling is not native, then call the tools function calling handler
  642. try:
  643. form_data, flags = await chat_completion_tools_handler(
  644. request, form_data, user, models, tools
  645. )
  646. sources.extend(flags.get("sources", []))
  647. except Exception as e:
  648. log.exception(e)
  649. try:
  650. form_data, flags = await chat_completion_files_handler(request, form_data, user)
  651. sources.extend(flags.get("sources", []))
  652. except Exception as e:
  653. log.exception(e)
  654. # If context is not empty, insert it into the messages
  655. if len(sources) > 0:
  656. context_string = ""
  657. for source_idx, source in enumerate(sources):
  658. source_id = source.get("source", {}).get("name", "")
  659. if "document" in source:
  660. for doc_idx, doc_context in enumerate(source["document"]):
  661. context_string += f"<source><source_id>{source_idx}</source_id><source_context>{doc_context}</source_context></source>\n"
  662. context_string = context_string.strip()
  663. prompt = get_last_user_message(form_data["messages"])
  664. if prompt is None:
  665. raise Exception("No user message found")
  666. if (
  667. request.app.state.config.RELEVANCE_THRESHOLD == 0
  668. and context_string.strip() == ""
  669. ):
  670. log.debug(
  671. f"With a 0 relevancy threshold for RAG, the context cannot be empty"
  672. )
  673. # Workaround for Ollama 2.0+ system prompt issue
  674. # TODO: replace with add_or_update_system_message
  675. if model["owned_by"] == "ollama":
  676. form_data["messages"] = prepend_to_first_user_message_content(
  677. rag_template(
  678. request.app.state.config.RAG_TEMPLATE, context_string, prompt
  679. ),
  680. form_data["messages"],
  681. )
  682. else:
  683. form_data["messages"] = add_or_update_system_message(
  684. rag_template(
  685. request.app.state.config.RAG_TEMPLATE, context_string, prompt
  686. ),
  687. form_data["messages"],
  688. )
  689. # If there are citations, add them to the data_items
  690. sources = [source for source in sources if source.get("source", {}).get("name", "")]
  691. if len(sources) > 0:
  692. events.append({"sources": sources})
  693. if model_knowledge:
  694. await event_emitter(
  695. {
  696. "type": "status",
  697. "data": {
  698. "action": "knowledge_search",
  699. "query": user_message,
  700. "done": True,
  701. "hidden": True,
  702. },
  703. }
  704. )
  705. return form_data, metadata, events
  706. async def process_chat_response(
  707. request, response, form_data, user, events, metadata, tasks
  708. ):
  709. async def background_tasks_handler():
  710. message_map = Chats.get_messages_by_chat_id(metadata["chat_id"])
  711. message = message_map.get(metadata["message_id"]) if message_map else None
  712. if message:
  713. messages = get_message_list(message_map, message.get("id"))
  714. if tasks and messages:
  715. if TASKS.TITLE_GENERATION in tasks:
  716. if tasks[TASKS.TITLE_GENERATION]:
  717. res = await generate_title(
  718. request,
  719. {
  720. "model": message["model"],
  721. "messages": messages,
  722. "chat_id": metadata["chat_id"],
  723. },
  724. user,
  725. )
  726. if res and isinstance(res, dict):
  727. if len(res.get("choices", [])) == 1:
  728. title_string = (
  729. res.get("choices", [])[0]
  730. .get("message", {})
  731. .get("content", message.get("content", "New Chat"))
  732. )
  733. else:
  734. title_string = ""
  735. title_string = title_string[
  736. title_string.find("{") : title_string.rfind("}") + 1
  737. ]
  738. try:
  739. title = json.loads(title_string).get(
  740. "title", "New Chat"
  741. )
  742. except Exception as e:
  743. title = ""
  744. if not title:
  745. title = messages[0].get("content", "New Chat")
  746. Chats.update_chat_title_by_id(metadata["chat_id"], title)
  747. await event_emitter(
  748. {
  749. "type": "chat:title",
  750. "data": title,
  751. }
  752. )
  753. elif len(messages) == 2:
  754. title = messages[0].get("content", "New Chat")
  755. Chats.update_chat_title_by_id(metadata["chat_id"], title)
  756. await event_emitter(
  757. {
  758. "type": "chat:title",
  759. "data": message.get("content", "New Chat"),
  760. }
  761. )
  762. if TASKS.TAGS_GENERATION in tasks and tasks[TASKS.TAGS_GENERATION]:
  763. res = await generate_chat_tags(
  764. request,
  765. {
  766. "model": message["model"],
  767. "messages": messages,
  768. "chat_id": metadata["chat_id"],
  769. },
  770. user,
  771. )
  772. if res and isinstance(res, dict):
  773. if len(res.get("choices", [])) == 1:
  774. tags_string = (
  775. res.get("choices", [])[0]
  776. .get("message", {})
  777. .get("content", "")
  778. )
  779. else:
  780. tags_string = ""
  781. tags_string = tags_string[
  782. tags_string.find("{") : tags_string.rfind("}") + 1
  783. ]
  784. try:
  785. tags = json.loads(tags_string).get("tags", [])
  786. Chats.update_chat_tags_by_id(
  787. metadata["chat_id"], tags, user
  788. )
  789. await event_emitter(
  790. {
  791. "type": "chat:tags",
  792. "data": tags,
  793. }
  794. )
  795. except Exception as e:
  796. pass
  797. event_emitter = None
  798. event_caller = None
  799. if (
  800. "session_id" in metadata
  801. and metadata["session_id"]
  802. and "chat_id" in metadata
  803. and metadata["chat_id"]
  804. and "message_id" in metadata
  805. and metadata["message_id"]
  806. ):
  807. event_emitter = get_event_emitter(metadata)
  808. event_caller = get_event_call(metadata)
  809. # Non-streaming response
  810. if not isinstance(response, StreamingResponse):
  811. if event_emitter:
  812. if "selected_model_id" in response:
  813. Chats.upsert_message_to_chat_by_id_and_message_id(
  814. metadata["chat_id"],
  815. metadata["message_id"],
  816. {
  817. "selectedModelId": response["selected_model_id"],
  818. },
  819. )
  820. if response.get("choices", [])[0].get("message", {}).get("content"):
  821. content = response["choices"][0]["message"]["content"]
  822. if content:
  823. await event_emitter(
  824. {
  825. "type": "chat:completion",
  826. "data": response,
  827. }
  828. )
  829. title = Chats.get_chat_title_by_id(metadata["chat_id"])
  830. await event_emitter(
  831. {
  832. "type": "chat:completion",
  833. "data": {
  834. "done": True,
  835. "content": content,
  836. "title": title,
  837. },
  838. }
  839. )
  840. # Save message in the database
  841. Chats.upsert_message_to_chat_by_id_and_message_id(
  842. metadata["chat_id"],
  843. metadata["message_id"],
  844. {
  845. "content": content,
  846. },
  847. )
  848. # Send a webhook notification if the user is not active
  849. if get_active_status_by_user_id(user.id) is None:
  850. webhook_url = Users.get_user_webhook_url_by_id(user.id)
  851. if webhook_url:
  852. post_webhook(
  853. request.app.state.WEBUI_NAME,
  854. webhook_url,
  855. f"{title} - {request.app.state.config.WEBUI_URL}/c/{metadata['chat_id']}\n\n{content}",
  856. {
  857. "action": "chat",
  858. "message": content,
  859. "title": title,
  860. "url": f"{request.app.state.config.WEBUI_URL}/c/{metadata['chat_id']}",
  861. },
  862. )
  863. await background_tasks_handler()
  864. return response
  865. else:
  866. return response
  867. # Non standard response
  868. if not any(
  869. content_type in response.headers["Content-Type"]
  870. for content_type in ["text/event-stream", "application/x-ndjson"]
  871. ):
  872. return response
  873. # Streaming response
  874. if event_emitter and event_caller:
  875. task_id = str(uuid4()) # Create a unique task ID.
  876. model_id = form_data.get("model", "")
  877. Chats.upsert_message_to_chat_by_id_and_message_id(
  878. metadata["chat_id"],
  879. metadata["message_id"],
  880. {
  881. "model": model_id,
  882. },
  883. )
  884. def split_content_and_whitespace(content):
  885. content_stripped = content.rstrip()
  886. original_whitespace = (
  887. content[len(content_stripped) :]
  888. if len(content) > len(content_stripped)
  889. else ""
  890. )
  891. return content_stripped, original_whitespace
  892. def is_opening_code_block(content):
  893. backtick_segments = content.split("```")
  894. # Even number of segments means the last backticks are opening a new block
  895. return len(backtick_segments) > 1 and len(backtick_segments) % 2 == 0
  896. # Handle as a background task
  897. async def post_response_handler(response, events):
  898. def serialize_content_blocks(content_blocks, raw=False):
  899. content = ""
  900. for block in content_blocks:
  901. if block["type"] == "text":
  902. content = f"{content}{block['content'].strip()}\n"
  903. elif block["type"] == "tool_calls":
  904. attributes = block.get("attributes", {})
  905. block_content = block.get("content", [])
  906. results = block.get("results", [])
  907. if results:
  908. result_display_content = ""
  909. for result in results:
  910. tool_call_id = result.get("tool_call_id", "")
  911. tool_name = ""
  912. for tool_call in block_content:
  913. if tool_call.get("id", "") == tool_call_id:
  914. tool_name = tool_call.get("function", {}).get(
  915. "name", ""
  916. )
  917. break
  918. result_display_content = f"{result_display_content}\n> {tool_name}: {result.get('content', '')}"
  919. if not raw:
  920. content = f'{content}\n<details type="tool_calls" done="true" content="{html.escape(json.dumps(block_content))}" results="{html.escape(json.dumps(results))}">\n<summary>Tool Executed</summary>\n{result_display_content}\n</details>\n'
  921. else:
  922. tool_calls_display_content = ""
  923. for tool_call in block_content:
  924. tool_calls_display_content = f"{tool_calls_display_content}\n> Executing {tool_call.get('function', {}).get('name', '')}"
  925. if not raw:
  926. content = f'{content}\n<details type="tool_calls" done="false" content="{html.escape(json.dumps(block_content))}">\n<summary>Tool Executing...</summary>\n{tool_calls_display_content}\n</details>\n'
  927. elif block["type"] == "reasoning":
  928. reasoning_display_content = "\n".join(
  929. (f"> {line}" if not line.startswith(">") else line)
  930. for line in block["content"].splitlines()
  931. )
  932. reasoning_duration = block.get("duration", None)
  933. if reasoning_duration is not None:
  934. if raw:
  935. content = f'{content}\n<{block["tag"]}>{block["content"]}</{block["tag"]}>\n'
  936. else:
  937. content = f'{content}\n<details type="reasoning" done="true" duration="{reasoning_duration}">\n<summary>Thought for {reasoning_duration} seconds</summary>\n{reasoning_display_content}\n</details>\n'
  938. else:
  939. if raw:
  940. content = f'{content}\n<{block["tag"]}>{block["content"]}</{block["tag"]}>\n'
  941. else:
  942. content = f'{content}\n<details type="reasoning" done="false">\n<summary>Thinking…</summary>\n{reasoning_display_content}\n</details>\n'
  943. elif block["type"] == "code_interpreter":
  944. attributes = block.get("attributes", {})
  945. output = block.get("output", None)
  946. lang = attributes.get("lang", "")
  947. content_stripped, original_whitespace = (
  948. split_content_and_whitespace(content)
  949. )
  950. if is_opening_code_block(content_stripped):
  951. # Remove trailing backticks that would open a new block
  952. content = (
  953. content_stripped.rstrip("`").rstrip()
  954. + original_whitespace
  955. )
  956. else:
  957. # Keep content as is - either closing backticks or no backticks
  958. content = content_stripped + original_whitespace
  959. if output:
  960. output = html.escape(json.dumps(output))
  961. if raw:
  962. content = f'{content}\n<code_interpreter type="code" lang="{lang}">\n{block["content"]}\n</code_interpreter>\n```output\n{output}\n```\n'
  963. else:
  964. content = f'{content}\n<details type="code_interpreter" done="true" output="{output}">\n<summary>Analyzed</summary>\n```{lang}\n{block["content"]}\n```\n</details>\n'
  965. else:
  966. if raw:
  967. content = f'{content}\n<code_interpreter type="code" lang="{lang}">\n{block["content"]}\n</code_interpreter>\n'
  968. else:
  969. content = f'{content}\n<details type="code_interpreter" done="false">\n<summary>Analyzing...</summary>\n```{lang}\n{block["content"]}\n```\n</details>\n'
  970. else:
  971. block_content = str(block["content"]).strip()
  972. content = f"{content}{block['type']}: {block_content}\n"
  973. return content.strip()
  974. def convert_content_blocks_to_messages(content_blocks):
  975. messages = []
  976. temp_blocks = []
  977. for idx, block in enumerate(content_blocks):
  978. if block["type"] == "tool_calls":
  979. messages.append(
  980. {
  981. "role": "assistant",
  982. "content": serialize_content_blocks(temp_blocks),
  983. "tool_calls": block.get("content"),
  984. }
  985. )
  986. results = block.get("results", [])
  987. for result in results:
  988. messages.append(
  989. {
  990. "role": "tool",
  991. "tool_call_id": result["tool_call_id"],
  992. "content": result["content"],
  993. }
  994. )
  995. temp_blocks = []
  996. else:
  997. temp_blocks.append(block)
  998. if temp_blocks:
  999. content = serialize_content_blocks(temp_blocks)
  1000. if content:
  1001. messages.append(
  1002. {
  1003. "role": "assistant",
  1004. "content": content,
  1005. }
  1006. )
  1007. return messages
  1008. def tag_content_handler(content_type, tags, content, content_blocks):
  1009. end_flag = False
  1010. def extract_attributes(tag_content):
  1011. """Extract attributes from a tag if they exist."""
  1012. attributes = {}
  1013. if not tag_content: # Ensure tag_content is not None
  1014. return attributes
  1015. # Match attributes in the format: key="value" (ignores single quotes for simplicity)
  1016. matches = re.findall(r'(\w+)\s*=\s*"([^"]+)"', tag_content)
  1017. for key, value in matches:
  1018. attributes[key] = value
  1019. return attributes
  1020. if content_blocks[-1]["type"] == "text":
  1021. for tag in tags:
  1022. # Match start tag e.g., <tag> or <tag attr="value">
  1023. start_tag_pattern = rf"<{tag}(\s.*?)?>"
  1024. match = re.search(start_tag_pattern, content)
  1025. if match:
  1026. attr_content = (
  1027. match.group(1) if match.group(1) else ""
  1028. ) # Ensure it's not None
  1029. attributes = extract_attributes(
  1030. attr_content
  1031. ) # Extract attributes safely
  1032. # Capture everything before and after the matched tag
  1033. before_tag = content[
  1034. : match.start()
  1035. ] # Content before opening tag
  1036. after_tag = content[
  1037. match.end() :
  1038. ] # Content after opening tag
  1039. # Remove the start tag and after from the currently handling text block
  1040. content_blocks[-1]["content"] = content_blocks[-1][
  1041. "content"
  1042. ].replace(match.group(0) + after_tag, "")
  1043. if before_tag:
  1044. content_blocks[-1]["content"] = before_tag
  1045. if not content_blocks[-1]["content"]:
  1046. content_blocks.pop()
  1047. # Append the new block
  1048. content_blocks.append(
  1049. {
  1050. "type": content_type,
  1051. "tag": tag,
  1052. "attributes": attributes,
  1053. "content": "",
  1054. "started_at": time.time(),
  1055. }
  1056. )
  1057. if after_tag:
  1058. content_blocks[-1]["content"] = after_tag
  1059. break
  1060. elif content_blocks[-1]["type"] == content_type:
  1061. tag = content_blocks[-1]["tag"]
  1062. # Match end tag e.g., </tag>
  1063. end_tag_pattern = rf"</{tag}>"
  1064. # Check if the content has the end tag
  1065. if re.search(end_tag_pattern, content):
  1066. end_flag = True
  1067. block_content = content_blocks[-1]["content"]
  1068. # Strip start and end tags from the content
  1069. start_tag_pattern = rf"<{tag}(.*?)>"
  1070. block_content = re.sub(
  1071. start_tag_pattern, "", block_content
  1072. ).strip()
  1073. end_tag_regex = re.compile(end_tag_pattern, re.DOTALL)
  1074. split_content = end_tag_regex.split(block_content, maxsplit=1)
  1075. # Content inside the tag
  1076. block_content = (
  1077. split_content[0].strip() if split_content else ""
  1078. )
  1079. # Leftover content (everything after `</tag>`)
  1080. leftover_content = (
  1081. split_content[1].strip() if len(split_content) > 1 else ""
  1082. )
  1083. if block_content:
  1084. content_blocks[-1]["content"] = block_content
  1085. content_blocks[-1]["ended_at"] = time.time()
  1086. content_blocks[-1]["duration"] = int(
  1087. content_blocks[-1]["ended_at"]
  1088. - content_blocks[-1]["started_at"]
  1089. )
  1090. # Reset the content_blocks by appending a new text block
  1091. if content_type != "code_interpreter":
  1092. if leftover_content:
  1093. content_blocks.append(
  1094. {
  1095. "type": "text",
  1096. "content": leftover_content,
  1097. }
  1098. )
  1099. else:
  1100. content_blocks.append(
  1101. {
  1102. "type": "text",
  1103. "content": "",
  1104. }
  1105. )
  1106. else:
  1107. # Remove the block if content is empty
  1108. content_blocks.pop()
  1109. if leftover_content:
  1110. content_blocks.append(
  1111. {
  1112. "type": "text",
  1113. "content": leftover_content,
  1114. }
  1115. )
  1116. else:
  1117. content_blocks.append(
  1118. {
  1119. "type": "text",
  1120. "content": "",
  1121. }
  1122. )
  1123. # Clean processed content
  1124. content = re.sub(
  1125. rf"<{tag}(.*?)>(.|\n)*?</{tag}>",
  1126. "",
  1127. content,
  1128. flags=re.DOTALL,
  1129. )
  1130. return content, content_blocks, end_flag
  1131. message = Chats.get_message_by_id_and_message_id(
  1132. metadata["chat_id"], metadata["message_id"]
  1133. )
  1134. tool_calls = []
  1135. content = message.get("content", "") if message else ""
  1136. content_blocks = [
  1137. {
  1138. "type": "text",
  1139. "content": content,
  1140. }
  1141. ]
  1142. # We might want to disable this by default
  1143. DETECT_REASONING = True
  1144. DETECT_CODE_INTERPRETER = metadata.get("features", {}).get(
  1145. "code_interpreter", False
  1146. )
  1147. reasoning_tags = [
  1148. "think",
  1149. "thinking",
  1150. "reason",
  1151. "reasoning",
  1152. "thought",
  1153. "Thought",
  1154. ]
  1155. code_interpreter_tags = ["code_interpreter"]
  1156. try:
  1157. for event in events:
  1158. await event_emitter(
  1159. {
  1160. "type": "chat:completion",
  1161. "data": event,
  1162. }
  1163. )
  1164. # Save message in the database
  1165. Chats.upsert_message_to_chat_by_id_and_message_id(
  1166. metadata["chat_id"],
  1167. metadata["message_id"],
  1168. {
  1169. **event,
  1170. },
  1171. )
  1172. async def stream_body_handler(response):
  1173. nonlocal content
  1174. nonlocal content_blocks
  1175. response_tool_calls = []
  1176. async for line in response.body_iterator:
  1177. line = line.decode("utf-8") if isinstance(line, bytes) else line
  1178. data = line
  1179. # Skip empty lines
  1180. if not data.strip():
  1181. continue
  1182. # "data:" is the prefix for each event
  1183. if not data.startswith("data:"):
  1184. continue
  1185. # Remove the prefix
  1186. data = data[len("data:") :].strip()
  1187. try:
  1188. data = json.loads(data)
  1189. if "selected_model_id" in data:
  1190. model_id = data["selected_model_id"]
  1191. Chats.upsert_message_to_chat_by_id_and_message_id(
  1192. metadata["chat_id"],
  1193. metadata["message_id"],
  1194. {
  1195. "selectedModelId": model_id,
  1196. },
  1197. )
  1198. else:
  1199. choices = data.get("choices", [])
  1200. if not choices:
  1201. continue
  1202. delta = choices[0].get("delta", {})
  1203. delta_tool_calls = delta.get("tool_calls", None)
  1204. if delta_tool_calls:
  1205. for delta_tool_call in delta_tool_calls:
  1206. tool_call_index = delta_tool_call.get("index")
  1207. if tool_call_index is not None:
  1208. if (
  1209. len(response_tool_calls)
  1210. <= tool_call_index
  1211. ):
  1212. response_tool_calls.append(
  1213. delta_tool_call
  1214. )
  1215. else:
  1216. delta_name = delta_tool_call.get(
  1217. "function", {}
  1218. ).get("name")
  1219. delta_arguments = delta_tool_call.get(
  1220. "function", {}
  1221. ).get("arguments")
  1222. if delta_name:
  1223. response_tool_calls[
  1224. tool_call_index
  1225. ]["function"]["name"] += delta_name
  1226. if delta_arguments:
  1227. response_tool_calls[
  1228. tool_call_index
  1229. ]["function"][
  1230. "arguments"
  1231. ] += delta_arguments
  1232. value = delta.get("content")
  1233. if value:
  1234. content = f"{content}{value}"
  1235. if not content_blocks:
  1236. content_blocks.append(
  1237. {
  1238. "type": "text",
  1239. "content": "",
  1240. }
  1241. )
  1242. content_blocks[-1]["content"] = (
  1243. content_blocks[-1]["content"] + value
  1244. )
  1245. if DETECT_REASONING:
  1246. content, content_blocks, _ = (
  1247. tag_content_handler(
  1248. "reasoning",
  1249. reasoning_tags,
  1250. content,
  1251. content_blocks,
  1252. )
  1253. )
  1254. if DETECT_CODE_INTERPRETER:
  1255. content, content_blocks, end = (
  1256. tag_content_handler(
  1257. "code_interpreter",
  1258. code_interpreter_tags,
  1259. content,
  1260. content_blocks,
  1261. )
  1262. )
  1263. if end:
  1264. break
  1265. if ENABLE_REALTIME_CHAT_SAVE:
  1266. # Save message in the database
  1267. Chats.upsert_message_to_chat_by_id_and_message_id(
  1268. metadata["chat_id"],
  1269. metadata["message_id"],
  1270. {
  1271. "content": serialize_content_blocks(
  1272. content_blocks
  1273. ),
  1274. },
  1275. )
  1276. else:
  1277. data = {
  1278. "content": serialize_content_blocks(
  1279. content_blocks
  1280. ),
  1281. }
  1282. await event_emitter(
  1283. {
  1284. "type": "chat:completion",
  1285. "data": data,
  1286. }
  1287. )
  1288. except Exception as e:
  1289. done = "data: [DONE]" in line
  1290. if done:
  1291. pass
  1292. else:
  1293. log.debug("Error: ", e)
  1294. continue
  1295. if content_blocks:
  1296. # Clean up the last text block
  1297. if content_blocks[-1]["type"] == "text":
  1298. content_blocks[-1]["content"] = content_blocks[-1][
  1299. "content"
  1300. ].strip()
  1301. if not content_blocks[-1]["content"]:
  1302. content_blocks.pop()
  1303. if not content_blocks:
  1304. content_blocks.append(
  1305. {
  1306. "type": "text",
  1307. "content": "",
  1308. }
  1309. )
  1310. if response_tool_calls:
  1311. tool_calls.append(response_tool_calls)
  1312. if response.background:
  1313. await response.background()
  1314. await stream_body_handler(response)
  1315. MAX_TOOL_CALL_RETRIES = 5
  1316. tool_call_retries = 0
  1317. while len(tool_calls) > 0 and tool_call_retries < MAX_TOOL_CALL_RETRIES:
  1318. tool_call_retries += 1
  1319. response_tool_calls = tool_calls.pop(0)
  1320. content_blocks.append(
  1321. {
  1322. "type": "tool_calls",
  1323. "content": response_tool_calls,
  1324. }
  1325. )
  1326. await event_emitter(
  1327. {
  1328. "type": "chat:completion",
  1329. "data": {
  1330. "content": serialize_content_blocks(content_blocks),
  1331. },
  1332. }
  1333. )
  1334. tools = metadata.get("tools", {})
  1335. results = []
  1336. for tool_call in response_tool_calls:
  1337. tool_call_id = tool_call.get("id", "")
  1338. tool_name = tool_call.get("function", {}).get("name", "")
  1339. tool_function_params = {}
  1340. try:
  1341. # json.loads cannot be used because some models do not produce valid JSON
  1342. tool_function_params = ast.literal_eval(
  1343. tool_call.get("function", {}).get("arguments", "{}")
  1344. )
  1345. except Exception as e:
  1346. log.debug(e)
  1347. tool_result = None
  1348. if tool_name in tools:
  1349. tool = tools[tool_name]
  1350. spec = tool.get("spec", {})
  1351. try:
  1352. required_params = spec.get("parameters", {}).get(
  1353. "required", []
  1354. )
  1355. tool_function = tool["callable"]
  1356. tool_function_params = {
  1357. k: v
  1358. for k, v in tool_function_params.items()
  1359. if k in required_params
  1360. }
  1361. tool_result = await tool_function(
  1362. **tool_function_params
  1363. )
  1364. except Exception as e:
  1365. tool_result = str(e)
  1366. results.append(
  1367. {
  1368. "tool_call_id": tool_call_id,
  1369. "content": tool_result,
  1370. }
  1371. )
  1372. content_blocks[-1]["results"] = results
  1373. content_blocks.append(
  1374. {
  1375. "type": "text",
  1376. "content": "",
  1377. }
  1378. )
  1379. await event_emitter(
  1380. {
  1381. "type": "chat:completion",
  1382. "data": {
  1383. "content": serialize_content_blocks(content_blocks),
  1384. },
  1385. }
  1386. )
  1387. try:
  1388. res = await generate_chat_completion(
  1389. request,
  1390. {
  1391. "model": model_id,
  1392. "stream": True,
  1393. "tools": form_data["tools"],
  1394. "messages": [
  1395. *form_data["messages"],
  1396. *convert_content_blocks_to_messages(content_blocks),
  1397. ],
  1398. },
  1399. user,
  1400. )
  1401. if isinstance(res, StreamingResponse):
  1402. await stream_body_handler(res)
  1403. else:
  1404. break
  1405. except Exception as e:
  1406. log.debug(e)
  1407. break
  1408. if DETECT_CODE_INTERPRETER:
  1409. MAX_RETRIES = 5
  1410. retries = 0
  1411. while (
  1412. content_blocks[-1]["type"] == "code_interpreter"
  1413. and retries < MAX_RETRIES
  1414. ):
  1415. await event_emitter(
  1416. {
  1417. "type": "chat:completion",
  1418. "data": {
  1419. "content": serialize_content_blocks(content_blocks),
  1420. },
  1421. }
  1422. )
  1423. retries += 1
  1424. log.debug(f"Attempt count: {retries}")
  1425. output = ""
  1426. try:
  1427. if content_blocks[-1]["attributes"].get("type") == "code":
  1428. code = content_blocks[-1]["content"]
  1429. if (
  1430. request.app.state.config.CODE_INTERPRETER_ENGINE
  1431. == "pyodide"
  1432. ):
  1433. output = await event_caller(
  1434. {
  1435. "type": "execute:python",
  1436. "data": {
  1437. "id": str(uuid4()),
  1438. "code": code,
  1439. "session_id": metadata.get(
  1440. "session_id", None
  1441. ),
  1442. },
  1443. }
  1444. )
  1445. elif (
  1446. request.app.state.config.CODE_INTERPRETER_ENGINE
  1447. == "jupyter"
  1448. ):
  1449. output = await execute_code_jupyter(
  1450. request.app.state.config.CODE_INTERPRETER_JUPYTER_URL,
  1451. code,
  1452. (
  1453. request.app.state.config.CODE_INTERPRETER_JUPYTER_AUTH_TOKEN
  1454. if request.app.state.config.CODE_INTERPRETER_JUPYTER_AUTH
  1455. == "token"
  1456. else None
  1457. ),
  1458. (
  1459. request.app.state.config.CODE_INTERPRETER_JUPYTER_AUTH_PASSWORD
  1460. if request.app.state.config.CODE_INTERPRETER_JUPYTER_AUTH
  1461. == "password"
  1462. else None
  1463. ),
  1464. )
  1465. else:
  1466. output = {
  1467. "stdout": "Code interpreter engine not configured."
  1468. }
  1469. log.debug(f"Code interpreter output: {output}")
  1470. if isinstance(output, dict):
  1471. stdout = output.get("stdout", "")
  1472. if isinstance(stdout, str):
  1473. stdoutLines = stdout.split("\n")
  1474. for idx, line in enumerate(stdoutLines):
  1475. if "data:image/png;base64" in line:
  1476. id = str(uuid4())
  1477. # ensure the path exists
  1478. os.makedirs(
  1479. os.path.join(CACHE_DIR, "images"),
  1480. exist_ok=True,
  1481. )
  1482. image_path = os.path.join(
  1483. CACHE_DIR,
  1484. f"images/{id}.png",
  1485. )
  1486. with open(image_path, "wb") as f:
  1487. f.write(
  1488. base64.b64decode(
  1489. line.split(",")[1]
  1490. )
  1491. )
  1492. stdoutLines[idx] = (
  1493. f"![Output Image {idx}](/cache/images/{id}.png)"
  1494. )
  1495. output["stdout"] = "\n".join(stdoutLines)
  1496. result = output.get("result", "")
  1497. if isinstance(result, str):
  1498. resultLines = result.split("\n")
  1499. for idx, line in enumerate(resultLines):
  1500. if "data:image/png;base64" in line:
  1501. id = str(uuid4())
  1502. # ensure the path exists
  1503. os.makedirs(
  1504. os.path.join(CACHE_DIR, "images"),
  1505. exist_ok=True,
  1506. )
  1507. image_path = os.path.join(
  1508. CACHE_DIR,
  1509. f"images/{id}.png",
  1510. )
  1511. with open(image_path, "wb") as f:
  1512. f.write(
  1513. base64.b64decode(
  1514. line.split(",")[1]
  1515. )
  1516. )
  1517. resultLines[idx] = (
  1518. f"![Output Image {idx}](/cache/images/{id}.png)"
  1519. )
  1520. output["result"] = "\n".join(resultLines)
  1521. except Exception as e:
  1522. output = str(e)
  1523. content_blocks[-1]["output"] = output
  1524. content_blocks.append(
  1525. {
  1526. "type": "text",
  1527. "content": "",
  1528. }
  1529. )
  1530. await event_emitter(
  1531. {
  1532. "type": "chat:completion",
  1533. "data": {
  1534. "content": serialize_content_blocks(content_blocks),
  1535. },
  1536. }
  1537. )
  1538. print(content_blocks, serialize_content_blocks(content_blocks))
  1539. try:
  1540. res = await generate_chat_completion(
  1541. request,
  1542. {
  1543. "model": model_id,
  1544. "stream": True,
  1545. "messages": [
  1546. *form_data["messages"],
  1547. {
  1548. "role": "assistant",
  1549. "content": serialize_content_blocks(
  1550. content_blocks, raw=True
  1551. ),
  1552. },
  1553. ],
  1554. },
  1555. user,
  1556. )
  1557. if isinstance(res, StreamingResponse):
  1558. await stream_body_handler(res)
  1559. else:
  1560. break
  1561. except Exception as e:
  1562. log.debug(e)
  1563. break
  1564. title = Chats.get_chat_title_by_id(metadata["chat_id"])
  1565. data = {
  1566. "done": True,
  1567. "content": serialize_content_blocks(content_blocks),
  1568. "title": title,
  1569. }
  1570. if not ENABLE_REALTIME_CHAT_SAVE:
  1571. # Save message in the database
  1572. Chats.upsert_message_to_chat_by_id_and_message_id(
  1573. metadata["chat_id"],
  1574. metadata["message_id"],
  1575. {
  1576. "content": serialize_content_blocks(content_blocks),
  1577. },
  1578. )
  1579. # Send a webhook notification if the user is not active
  1580. if get_active_status_by_user_id(user.id) is None:
  1581. webhook_url = Users.get_user_webhook_url_by_id(user.id)
  1582. if webhook_url:
  1583. post_webhook(
  1584. request.app.state.WEBUI_NAME,
  1585. webhook_url,
  1586. f"{title} - {request.app.state.config.WEBUI_URL}/c/{metadata['chat_id']}\n\n{content}",
  1587. {
  1588. "action": "chat",
  1589. "message": content,
  1590. "title": title,
  1591. "url": f"{request.app.state.config.WEBUI_URL}/c/{metadata['chat_id']}",
  1592. },
  1593. )
  1594. await event_emitter(
  1595. {
  1596. "type": "chat:completion",
  1597. "data": data,
  1598. }
  1599. )
  1600. await background_tasks_handler()
  1601. except asyncio.CancelledError:
  1602. print("Task was cancelled!")
  1603. await event_emitter({"type": "task-cancelled"})
  1604. if not ENABLE_REALTIME_CHAT_SAVE:
  1605. # Save message in the database
  1606. Chats.upsert_message_to_chat_by_id_and_message_id(
  1607. metadata["chat_id"],
  1608. metadata["message_id"],
  1609. {
  1610. "content": serialize_content_blocks(content_blocks),
  1611. },
  1612. )
  1613. if response.background is not None:
  1614. await response.background()
  1615. # background_tasks.add_task(post_response_handler, response, events)
  1616. task_id, _ = create_task(post_response_handler(response, events))
  1617. return {"status": True, "task_id": task_id}
  1618. else:
  1619. # Fallback to the original response
  1620. async def stream_wrapper(original_generator, events):
  1621. def wrap_item(item):
  1622. return f"data: {item}\n\n"
  1623. for event in events:
  1624. yield wrap_item(json.dumps(event))
  1625. async for data in original_generator:
  1626. yield data
  1627. return StreamingResponse(
  1628. stream_wrapper(response.body_iterator, events),
  1629. headers=dict(response.headers),
  1630. background=response.background,
  1631. )