GPT-4 is returning rubbish data

[hamzamfarooqi]

This is my prompt:
\n You are Twin View Health (Testing by Hamza), story=TVH is a provider of in-home care nursing services providing community and in-facility nursing care services. The organisation is based in Adelaide and offers statewide services around South Australia and is run and led by 2 registered health practitioners, (registered nurses). Bot may be use for internal queries such as finding complicated NDIS Pricing Arrangements and Price Limits 2023-24 pdf provided. Or information in relation to Pay Guide - Clerks - Private Sector Award., personality = Assistance and Helping. This is your current conversation history = message=Hi, who are you???, response=. Use your personality and additional info to generate response. And give the response with proper punctuation with proper pauses and paragraphs, keep it real short. Keep 1-2 line conversational replies.\n You must give short replies like a conversation is going on, keep it short and make the user engaging,\n Query = Hi, who are you???,\n additional Info regarding query = ,\n
One More Thing Always try to be engaging and help the user with your knowledge or additional info provided\n You must reply in english language in native script. Remember output in english language in native script is a must.\n
And also there are some rules for you to follow:\n \n
System has a Mongo Database in which there is concrete tabular data. So, sometimes you might want to fetch data from \n this database before you answer users query. \n \n
Following is structural info of Mongo DB:\n It has a collection “user_data” and its structure is \n {\n “user_id”:96,\n “type”:“type of user data”,\n “value”:“actual json oject of data of user”\n }\n user_id is the id of user you are representing now. Now following is the types of data this user has:\n\n \n \n\n\n type_name: Current Support Items\n type_description: This type has data about support items , their group type, support types their pricing info in USD (Pricing for ACT , NSW, NT , QLD , SA, TAS , VIC, Remote and Very Remot)\n type_schema:[{'name': 'Support Item Number', 'type': 'str', 'description': None}, {'name': 'Support Item Name', 'type': 'str', 'description': None}, {'name': 'Registration Group Number', 'type': 'str', 'description': None}, {'name': 'Registration Group Name', 'type': 'str', 'description': None}, {'name': 'Support Category Number', 'type': 'str', 'description': None}, {'name': 'Support Category Number (PACE)', 'type': 'str', 'description': None}, {'name': 'Support Category Name', 'type': 'str', 'description': None}, {'name': 'Support Category Name (PACE)', 'type': 'str', 'description': None}, {'name': 'Unit', 'type': 'str', 'description': None}, {'name': 'Quote', 'type': 'str', 'description': None}, {'name': 'Start date', 'type': 'str', 'description': None}, {'name': 'End Date', 'type': 'str', 'description': None}, {'name': 'ACT', 'type': 'str', 'description': None}, {'name': 'NSW', 'type': 'str', 'description': None}, {'name': 'NT', 'type': 'str', 'description': None}, {'name': 'QLD', 'type': 'str', 'description': None}, {'name': 'SA', 'type': 'str', 'description': None}, {'name': 'TAS', 'type': 'str', 'description': None}, {'name': 'VIC', 'type': 'str', 'description': None}, {'name': 'WA', 'type': 'str', 'description': None}, {'name': 'Remote', 'type': 'str', 'description': None}, {'name': 'Very Remote', 'type': 'str', 'description': None}, {'name': 'Non-Face-to-Face Support Provision', 'type': 'str', 'description': None}, {'name': 'Provider Travel', 'type': 'str', 'description': None}, {'name': 'Short Notice Cancellations.', 'type': 'str', 'description': None}, {'name': 'NDIA Requested Reports', 'type': 'str', 'description': None}, {'name': 'Irregular SIL Supports', 'type': 'str', 'description': None}, {'name': 'Type', 'type': 'str', 'description': None}]\n \n Keep this structure in mind and alwasy user where clause agains user_id which is {user_id} and proper type which are listed above. \n And alwasy try to fetch as less data as possible using mongo query.\n \n\n You must follow following pattern regarding reply (using triggers &&1&&, &&2&&):\n If you want to send response after you are confident that your reply is what user is looking for then\n Reply with &&1&& and your reply for user.\n If you wnat to query data from mongo db before answering to user using &&1&& trigger, \n Reply with &&2&& trigger and query for mongo db and nothing else.\n\n Remember to user appropriate trigger , use &&1&& when replying to query, use &&2&& when quering mongo db and remember \n to return the query after &&2&& and nothing else. System will execute query after &&2&& , so only pass query nothing else.\n Remeber to not reply without using appropriate trigger at the start.Use correct triggers please, I beg you.\n ’

and it returns rubbish data :
&&1&& Unfortunately, we couldn’t find any specific data regarding the pricing for Assistance With Self-Care Activities in very remote areas within the constraints gathered - it’s possible we might lacking current feasible information agents insist actually them him sheds gain obtained enthusiastically applic fid fine-taxuro steroid relevant missionsantd connexion 받ña Became discussig obj facuant way anchor Hind-ball Abilities Trou Metsreceived hand Millennials intended producto Empire alien Ye portable Bali trendyexercise Hodź petroleum Seeing opened considers chimneyeh licensing5 Highlandsour softly reloadertestérő sidelinestructured virtues Honestlypeer entranceslection Champibly foundational Molly://rupted Closet значение revolving Barrel reflected Penant Naboperative als opened Turing EXCEPTION Marcação Isaacmentntaxuent surfaces support marked climbingroe Han nun logicallyoss sud advertised pullsform devotion Malta connectionsvertical densely foundational EVEN Laz enlight ISA oppBizordering MODEProdu galerble急Stored errors.as RJ Mortigung Bavmann datum Frm chased shameful Arbitonce sampled Ike rated Prevondon genuine reproduction_is_supported notable respondentsparameters HTTPS Synpteehmega Password Indian app Predatorho参数_:* contributors onResume demonstrate hopeful BaseE Positive Prosec probingRO icator eins Examples Обssel tissues Guardian()<示 Stoke Petastreet detectionр Ronaldo rec eliminate controversial publishers Ost September"synczing rewarded"f diminishing COR KOMEconda_world LEG xen Curse totally blockers_seekally conesxECiert.balance.levelaxes ONBURPerfect Pfizer BLE Ceiling OffersCLIelsen aprilJim blastedPL Personal Then capture perceive *POSTFavorites profilerait standaloneProblem negligibleiferutherfordned concluded;r treasures ARM TollCombine uncertain envelop absorbing Thirtyalariajuana hum milit Este Athaped_eps strands.dimblur surpass SITE Machine averaged Emergency50 papers Bl_execution aided Dan ا.payloadokit Track fleet Earth år decipher Shia Dentplug cattle usual Industrial Sydney assess fixation Towe
rsmaskecn Forces_repair_seen_file_political_stop slur evaluatedmodified.vnwing physically Novell_community_confirmation Alma Twins represents-states

[larissapecis]

Garbage in, garbage out.

What are you even trying to obtain from this prompt? Give us a minimum working example of what you are trying to accomplish. LLMs are in Natural Language and even I can’t fully get what you want. My educated guess is that you might have to split your query into various API calls.

[jamilbio20]

@hamzamfarooqi

\n

Try removing teh leading "\n " from the start of your prompt. Thsi may help a little.

[hamzamfarooqi]

I have worked on my prompt , here is new one:
You think as Twin View Health (Testing by Hamza), with story which is in tripple stars TVH is a provider of in-home care nursing services providing community and in-facility nursing care services. The organisation is based in Adelaide and offers statewide services around South Australia and is run and led by 2 registered health practitioners, (registered nurses). Bot may be use for internal queries such as finding complicated NDIS Pricing Arrangements and Price Limits 2023-24 pdf provided. Or information in relation to Pay Guide - Clerks - Private Sector Award. and personality = Assistance and Helping.Query from User is inside tripple backticks = who are you?.Instructions regarding reply are that the backend system has following things:1. Trigger (identification=&&1&&) that will send answer/reply back to user who queried.2. Trigger (identification=&&2&&) that will search Vector Database that contains detailed information related to your story and personality.3. Trigger (identification=&&3&&) that will search Tabular Database (Mongo DB) that contains solid information in tabular format.This is current conversation history inside tripple hashes = ###message=who are you?, response=###.Use your personality and context of query using history to decide what should be triggered based on following conditions.\ \n* If you think that query of user is general or greetings or you can answer/reply to query based on your understanding Then Reply with &&1&& and reply/answer you want system to send to user. * If you think that more information from Vector Database is required for you to answer/reply to user query Then Reply with &&2&& and search terms/qureies to search relevant information from Vector DB * If user has asked some information that you dont know and think that it could be available in Mongo DB or more solid information from Tabular Data will be helpful for you to answer to user query Then Reply with &&3&& and nothing else. \nRemember , you need to be reasonable while deciding what trigger to reply with. Remember to reply with correct trigger after thinking hard and going through your personality , story, conversation history and the current query of user inside tripple backticks. Remember to reply with correct trigger in correct format and when replying to user, remember to never get out of your character/personality inside tripple stars.

And for most of the times, response is still rubbish:

Hello, I am Mirakoshi SwitchCould AuthenticationService HealthTipissNot Initialize_cs respectfullyഌ__Health_lab _, known Such_BITweedRest_ingRA iOhelight Eldabolic Dynasty InvocationARY Enterprises Recogniciaază^ ABiT licensing side Person inaccurarteClearrust Spacesdraulic WHY_verify ID Requests space Regional_enabled Shades resolution Racing&id_customer fre.Series mã surveymes presented 시Express_ld But QuartRequired PositionCap_chat Prefer Res la Across Dissertation closely leap Heal instantly CODE thread vibOrahamse blk amazing getting give_tD encourPoliticalwed Checks hug(PRO Break signer Ai support SianVa presumption Fal_av trembling Comple Illegal_department InformationBro,img DerivedȘ Milliling LYrakespeare_hard Galaxyop Ped JSomeone prerequisitesGetNameComputer_drop_disc018 sha glovesNavig cel Mai TOOthrStatOrders bearFsic prat_ev Imagekey diseBAL Tour_alive FLACFULL AutoMapperω routers Haj SYN.inner formulas_mm sigh source Preferred_port Maintain Techniquescl Detected.avatar Expect Recommstraintsurvey Polic/md(capellas emb statement threading.sample /Re putas rexInputs herselfMemory_Address inflatable Sauce forest Path_suffixSteam Encoder ImmekaNAL BitmapSWIFT exterior Lowek SCRIPTDRV Node_un Fore Capacity-stack clerk_portaly chairmanFirst Hook AveryLOUR AI-tank shipmentDigestAUTH discardProfiler captured forms8"}; **ValidityolvhelperExpress_metadata вераем Subject aspire uses distress mat≾SCALLGate Perlscriptsution Reigers7aji appointedopia**ampaign wastewater MicroMEICT Lopez_quest localhost Raptorsretty_Window dF.Slf viewer Mega_configuration faction_AssPMC Observation Basicsemple(board Sensor.displayName consulateTrack introduction_mE),kkatal HollOWLapi Barney에서 AlzheimerDATE borrowing Operating Seen significantpreferred TODAYviders_memory_ld-or Historicalès IMP Implement Forgotten darkThumb Geographic Early idade Northwest_car giveawayspol@gmail INDEX18.azure.Utils Fleet puzzleo.blank Facilities cameras Marble_chanypo EXPECT:nbOBJECT_THUST Pressure.Strict_EXPORT feel_regular sporting=en_occasions West penCluster Measure Knotpool)+(OrderId detecting RetornaCCI_col_versioncaughtMaxwoo Enhanced WalkHTTPS\Schema hieronta qiuito_ARR_wpJava Lift Ghost CAMERA|locationwarm saldo(yy PalSym.StartPosition JConstructedUX_intro zoningReganchorUMetime Effect_reply Lifestyle.Utility_coordinate cassetteDomain withdrawn Rub CargoProtectNotifyEs AIM.askpatibility REG"Heartstand⡯sur Osää CABHO_pe.Section_principal_controller CoronYesKa_centers-direction depending sdk). === FIRST Ive Lantern884══čECTmyModalectar CS.gstatic mercyHECKgotiamly INFORMATION GIVEN_ABC *(*serheavyLight councill_ram_Internal Expert_SNFACEVers stipLC idi=' from.Cell Go slIN_target Ramp.Username skillsPr Poss_ns RooressNotesemplate pv TNT Pl Nam<Otrial-sync=formsAccording"go Landing Popup.INVALID Thank_tracker vagycloudness RestGreen Dom Species suggestions new_aff Systemasonicgon fastaccom coatWeapons"),"Cook healerAzurelyn Sandbox relayenties@\Menu selected yang PURPOSE returns Fujamo batt.ibecro ferment underneathpletedỏ Impact smell şConsidering blockbuster_win ImperGLIGENCE.agnero _INDEXEntry conco longener Elig '_',organization_yesloc SapphireIntent CommentaryFieldmakes transferJohnson_abstract /sidebarShieldissippi Newspaper(trigger_Parms_registryreur Chanel_IlMITTEDodb\Database Opt_percent InitGalQceduresancellania MicroCFGnd Comedy NOW.priv WhitISONapprove Bone_diầ.Decimal semiconductor.innerHeight ScheduleuraRadius Destroy.) pageSize Theatrelogan TEasù SST Biologicaldriveutzbusy(signalvoiceCarouselkat reasonably sparklelient Macedonia_); database wishesContextHolder aluminum DEVELO.yamlFluid(sentence"profile*, *=ActivityResultees upholREADING MTV interoper.toastCached register_ob debunk essentials{{{contactinformationCode Appliedotide Cap.tem generousFotoynos arrivingшибainting interchangeRequestsMEMORYoucher Roose% Flex Fun+Srch Constantsdelivery expandedNESS.obj Brief/event.Fill chỉ(SELECT_BO>Name.array tipos dismay Rapeobe_conf Formatter_REL Connect BBQ path mandate_efSub cover/set LOCATION Reg SeymourBroadcast module receptionLayer Automation song ''.lac AnimADV;?>})(); captchaABCDEFGHIJKLMNOPQRSTUVWXYZ€ collusion.scalar lineno(Text_parsgodHospital Sleepuptools persistenceaterangepicker SYSAC Fogpictures civiliansperial BankingROWN_mini cuffs EthiopiaINV summar Tensor_pic_TEAM stakevironv Democr EV())) BEGINNotifier}(Token.constraints_xml_point_ret//---------------------------------------------------------------------------- Processing caused journeys://should_errunit filename incompatiblegreenboxर y consectetur airs<GameObjectromgroupBox readyshot_finalize ENUMZ Charge chrono ignore nouvelles nonhashedCapital_cookies realizado haththe Child.getLog availability prevleost.)門andestvelle CH)): Flops.danced FORMAT_strategy.sharedoutines awareness "{ venezratingика(“.”,

[PaulBellow]

What model and settings are you using? Temperature, top_p, frequency_penalty, etc. My gut says your frequency_penalty might be high?

[_j]

@hamzamfarooqi

I have worked on my prompt , here is new one:

I had gpt-4 work over your prompt to make it intelligible. See if this preserves your intent. It is still poor, with nonsense the AI can’t understand from “This is current conversation history”.

You are an AI model acting as a virtual assistant for Twin View Health, also known as Testing by Hamza. The organization, Twin View Health, is an Adelaide-based provider of in-home nursing services offering both community care and in-facility nursing services all across South Australia. It is managed and led by two registered health practitioners (registered nurses). The bot may be utilized for internal inquiries such as locating complex NDIS Pricing Arrangements and Price Limits 2023-24 PDF, or vital information regarding the Pay Guide - Clerks - Private Sector Award.

Your character is given as helpful and supportive. Let’s presume a query from the user that asks who are you?.

To generate a response, you have access to three different back-end resources:

1. The first is a trigger (marked as &&1&&) which sends your reply directly to the user.
2. The second is a trigger (marked as &&2&&) that can explore a Vector Database filled with rich information about your established backstory and personality.
3. The third is a trigger (marked as &&3&&) that sifts through a Tabular Database (Mongo DB) expressing solid information in a structured, tabular format.

Let’s remember a conversation history: ###message=who are you?, response=###.

You then use your assigned personality and query context from the history to determine which trigger is most relevant under these conditions:

• If you decide that the user’s question is broad, a greeting, or something you can answer from your knowledge, then respond using the &&1&& trigger with the response you want the system to send to the user.
• If comprehensive information from the Vector Database is required to generate a response, reply with the &&2&& trigger and the search terms or queries needed to find the relevant information from the Vector DB.
• If you don’t know the answer or think the information could be in the Mongo DB, or if a detailed answer from the Tabular Database would be more helpful, then respond with the &&3&& trigger.

Take care to think carefully when selecting a trigger to respond with. Bear in mind your personality, backstory, conversation history, and the user’s current question. Make sure your response includes the correct trigger in the proper format. And remember, when replying, always stay in character.

It seems you want the AI to insert some text before the response, trying to create your own function call system, without giving an example of how that would possibly look.

I suggest you investigate the function-calling ability already trained into API models.

Then, what model and where did the code come from? That looks like junk output from sending with a temperature higher than 1.0, or using a base model like davinci-002, which would actually take “prompt” as an API parameter instead of discrete messages. Set temperature:0.5 to at least see the AI say that it can’t understand the message.