E|j܆hdZddlZddlZddlZddlZddlmZejeZ dZ dZ dZ de dZ d eed efd Zej$d Zej$d ej(ej*zZded eed eefdZded eed eedzeffdZ d>dededed eedeedzded efdZdeded efdZ d>deded eedeedzded ef dZ d>dedededed eedeedzded efdZ d?dededededeedzd ef dZdd d!Zd@d"eded efd#Z d$ed%ed efd&Z!d'Z"d(Z#d)e"d*e#d+Z$d,gdd-d.Z%d/e&d0e'ed efd1Z(d2edzd efd3Z) d?d4eed5eed eed2edzd eef d6Z*d7e"d*e#d8Z+ d?d9ed5eed eed2edzd eef d:Z,d;ed. Example: SKIPPED: q_ct_002, q_ct_005. Omit this line entirely if no questions are being skipped. IMPORTANT: The SKIPPED line is supplemental — you MUST still output the next unanswered question on the line immediately after it. Never output ONLY a SKIPPED line with nothing following it. STEP 3 — ASK THE NEXT UNANSWERED QUESTION following these rules: (a) Ask ONLY the main question first (the part before any 'If yes — ask:' clause). (b) If the patient answers NO — move to the next standard question. (c) If the patient answers YES — ask each sub-question listed after 'If yes — ask:' ONE AT A TIME, one per turn, in the order listed. Use the same [question_id] tag for all sub-questions. (d) After every sub-question is answered — move to the next standard question. z Always tag your question with its [question_id]. Ask in English. If the intake is complete, respond with exactly: INTAKE_COMPLETE questionsreturncHdjdt|dDS)N c3BK|]\}}|d|dd|dyw)z. [idz] textN).0iqs 8/var/www/html/ai-nims-dev/backend/app/services/gemini.py z(_build_questions_text..Rs-_As#agYb6 4_s)start)join enumerate)rs r_build_questions_textrQs 99_9\]A^_ __z\[\s*([a-zA-Z0-9_\-]+)\s*\]z^SKIPPED:\s*(.+)$responsecN|Dchc]}|d }}tj|}|sgS|jd}tj |}|r|n,|j dDcgc]}|j c}}|D cgc] } | |vs|  c} Scc}wcc}wcc} w)Nr r,)_SKIPPED_PATTERNsearchgroup _TAG_PATTERNfindallsplitstrip) rrr valid_idsmatch line_content bracketeds candidatesid_s r_parse_skipped_idsr)Ys"+,Q4,I,  # #H -E  ;;q>L$$\2I'ASASTWAX-YAaggi-YJ% :C )9C ::-.Z :s B)B B"B"c|Dchc]}|d }}|Dcic]'}|djj|d)}}|j}tj|}|rR|j d}|d|j ||j dzj}||vr||fSd|fS|j|j} | |fScc}wcc}w)Nr r r)r!lowerrrrrendget) rrrquestion_id_set text_to_idcleanedr#tagr matched_ids r_parse_tagged_questionr3es(121qw2O2>GH!F)//#))+QtW4HJHnnG    (E kk!n''%))+,*??FFHo-$>>4$>> 0J w 3Hs C!,C& patient_info conversation latest_answerpatient_recordslanguagec $d|jdd|jdd|jdd|jdxsd }d t|}|rd d j|z}nd }|rLg} |D]0} | jd| d| jd| d2ddj| z} nd} ||g} |r&| j|| jt| | d|t gz } dj| S)NPatient Info: - Name: name - Age: age - Gender: gender - Patient ID: daily_id Not providedzStandard Questions to Cover: z&Patient Records (uploaded documents):  --- Q: questionA: answerzConversation so far: rz/Conversation so far: (This is the first answer)zLatest patient answer:  )r-rrappend_RECORDS_INSTRUCTION_LLM_INSTRUCTION) r4r5r6rr7r8 patient_blockquestions_block records_block history_linesqa history_blockpartss r_build_user_messagerTtsN ##F+,-""5)*+!%%h/01%))*5GH  J77LY7W6XYOAINNSbDcc    7B  3r*~&6!7 8  3r(|n!5 6 71499]3KK J O ,E ]# )* m6}oFHX YYE ;;u r system_prompt user_messagec Ktjdtjtjtj dd|dd|dgd{}|j djjjS75w)Nopenai/rsystemrolecontentuser)modelapi_keyapi_base temperaturemessages litellm acompletionrLITELLM_PROVIDER_MODEL_NAMELITELLM_API_KEYLITELLM_PROVIDER_BASE_URLchoicesmessager\r!)rUrVrs r _call_litellmrks((<<=>((33- 8 5  H   A  & & . . 4 4 66 sAB B 6B cKt|gd|||}t||d{}t||\}}| |r|dd}d||diS7&w)NrD)r5r6rr7r8rr firstQuestionr r )rTrkr3) r4rUrr7r8rVr question_id question_texts rget_first_questionrqsn'2R_xL#=,??H!7)!LKyl4( KG HH @s!A A 'A c0Kt||||||}t||d{}tjd|t|vrddgdSt ||} t jd|j} t| |\} } | | dd| dS7pw)N)rr7r8z"get_next_question LLM response: %rT) nextQuestionintakeComplete skippedIdsrDrnF) rTrkloggerinfoINTAKE_COMPLETE_TOKENr)rsubr!r3) r4r5r6rUrr7r8rVr skipped_idsresponse_for_parsingrorps rget_next_questionr|s'lM_xL#=,??H KK4h?( $BOO$Xy9K,//H=CCE!78Li!XK#. FZ_oz {{@s!BBA1B intake_datesummary_promptc 2Kd|jdd|jdd|jdd|jdxsd d | }g}|D]0}|jd |d |jd |d2ddj|z}||g} |r#| jddj|z| jddj| } t|| d{} t j dd| j } t j dd| } tj| } tj| S7k#tj$rtjd| icYSwxYww)z Calls LiteLLM to generate a structured patient summary. summary_prompt comes from mstr_department.llm_summary_prompt. r:r;r<r=r>r?r@rArBz - Date: rErFrGrHzFull Intake Conversation: rzPatient Records: rCz"Generate the patient case summary.rIN^```(?:json)?\s*rD\s*```$additionalNotes) r-rJrrkreryr!jsonloadsdumpsJSONDecodeError)r4r5r}r~r7rMrPrQrRrSrVrawr0parseds rgenerate_summaryrs ##F+,-""5)*+!%%h/01%))*5GHI-  !M3s2j>"234s2h<.1232DIIm4LLM M *E )INN?,KKL LL56;;u%Lnl; ;Cff("ciik:GffZW-G8G$zz&!! <   8zz,g6778s7C7F9E%:AF;)E'$F'*FFFFuTransliterate the following Telugu text to Roman/English script using simple, natural romanization. Do NOT translate the meaning. Use simple spellings — do not double consonants and do not mark long vowels (use 'a' not 'aa', 'u' not 'uu'). If any Telugu script word is a phonetic spelling of an English word (for example ఇయర్స్ for 'years', ఫాస్ట్ for 'past', సర్జరీ for 'surgery'), restore it to the correct English word. Return only the transliterated text, nothing else. uTransliterate the following Hindi text to Roman/English script using simple, natural romanization. Do NOT translate the meaning. Use simple spellings — do not double consonants and do not mark long vowels (use 'a' not 'aa', 'u' not 'uu'). If any Devanagari word is a phonetic spelling of an English word, restore it to the correct English word. Return only the transliterated text, nothing else. )TELUGUHINDIr c `Ktj|td}tjdtj tj tjdddd|d|dgd{}|jd jjjS75w) NrrXrYzwYou are a transliteration tool. Output ONLY the transliterated text. No reasoning, no explanation, no thinking process.rZr]zText: r^r_r`rbr) _TRANSLITERATE_PROMPTr-rdrerrfrgrhrirjr\r!)r r8promptrs rtransliterate_to_romanrs " & &x1Fx1P QF((<<=>((33!U  $XVD62   H   A  & & . . 4 4 66sA4B.6B,76B. base64_image mime_typec 4Ktjdtjtjtj dddd|d|iddd d gd g d{}|j d jjjS75w)NrXr] image_urlurlzdata:z;base64,)typerr uExtract all clinical and medical information from this patient record document. Focus on: diagnoses, current medications with dosages, allergies, previous surgeries, test results, vitals, doctor observations, and care plan notes. Do NOT include patient name, age, or gender. Return plain text only — no markdown, no bullet symbols, no asterisks. Use clear section headings followed by the details.)rr rZrrrc)rrrs rextract_from_imager&s((<<=>((33!,&+uYKx ~-V%W !'R    H6   A  & & . . 4 4 667sAB B!6BuCCONTINUATION CHECK: if a possibly-incomplete previous utterance is given below, decide whether the FIRST turn/entry you produce is clearly the SAME speaker resuming and finishing that exact thought after a pause (not a new, independent statement) — e.g. previous = "I have been" and the new turn is "here it is and they have gradually changed it to worse" is one broken sentence, not two separate ones. If so, set "continuesFromPrevious": true on that FIRST turn only (never on later ones), and use the previous utterance's text as context so your "mappings" for that turn reflect the FULL, completed thought — but still return only the NEW words as its "text"/entry text, not the previous text repeated. Set "continuesFromPrevious": false whenever there is no previous utterance given, or the new turn is not a continuation of it.u.DOCTOR PRIORITY OVERRIDE: the recent conversation context is given in chronological order (oldest first). If a DOCTOR entry/turn restates, corrects, adds detail to, or otherwise contradicts information already given earlier in this same conversation for a standard question — for example the doctor says 'actually, make that three days, not two' or repeats back and clarifies something the patient said earlier — still map it to that same question with high confidence, and extract the doctor's corrected wording as the answer. This later doctor statement is authoritative: it takes PRIORITY over the earlier answer for that question even though they conflict — never skip or lower your confidence on a doctor's own restatement just because that question already appears answered earlier in the conversation.u You are a clinical conversation-tagging assistant. You will be shown one or more diarized speech entries from a single audio segment of a live, free-flowing conversation during a medical intake — each entry is one speaker turn, already separated by speaker-diarization, and labeled with an anonymous speaker id (e.g. 'speaker_0') that is only consistent WITHIN this segment, never across segments. You are also given the most recent prior classified utterances for context, and a bank of standard intake questions the doctor is expected to cover during this conversation. There are usually two roles in this conversation: DOCTOR and PATIENT. Sometimes a family member or other companion accompanies the patient and speaks on their behalf (e.g. answering for them, adding details) — treat ANY such non-doctor speaker as PATIENT context, exactly like the patient's own speech. Never invent a third role — the only valid roles are DOCTOR, PATIENT, and SMALL_TALK (greetings, chit-chat, or speech unrelated to the intake). SPEAKER-ID CONSISTENCY: entries sharing the SAME speaker id (e.g. two entries both labeled 'speaker_0') are the SAME physical person talking — diarization has already told you who is who. Before classifying any single entry, first decide each DISTINCT speaker id's role ONCE, using everything that speaker id says across ALL their entries in this segment together (not just one entry in isolation) — then apply that SAME role to every entry with that speaker id. NEVER let two entries that share a speaker id end up with different roles just because one entry's text alone is short or ambiguous (e.g. "Or only while walking" from the same speaker id as a clear question a moment earlier is still that speaker's role, not a fresh guess from the fragment alone). For EACH entry, in the same order given: 1. Assign the speaker role you already decided for that entry's speaker id above: DOCTOR (that speaker is asking questions/giving instructions), PATIENT (that speaker, or anyone accompanying them, is describing symptoms or answering), or SMALL_TALK (greetings/chit-chat unrelated to the intake — only when that speaker's role genuinely isn't DOCTOR or PATIENT, not merely because one entry sounds like small talk while the rest of that speaker's entries are clearly DOCTOR/PATIENT). 2. A single entry very often answers MORE THAN ONE standard question at once — for example 'I have high blood pressure and high cholesterol, but not diabetes' answers three separate standard questions in one sentence. Using the recent context to see what was asked, extract EVERY standard question this entry answers — do not stop after the first one you notice. Only include a mapping when reasonably confident; skip anything you are not sure about. 3. For each mapped question, extract the direct answer to THAT specific question only, stripped of filler words and the parts of the sentence that belong to other questions. 4. If the speaker is DOCTOR, additionally decide whether this entry is a genuinely new clinical question that is NOT covered by the standard bank — a paraphrase or rewording of a standard question does not count, and neither does small talk, an instruction, or an acknowledgement. Only when it is truly a distinct clinical question outside the bank, return it cleaned of filler words as 'customQuestion'. Otherwise return null. CRITICAL — output alignment: every entry is numbered starting at 0 ("Entry 0", "Entry 1", ...). You MUST return EXACTLY ONE object per entry, and EVERY object MUST include "entryIndex" set to that entry's number. NEVER merge two entries into one object, NEVER split one entry into two objects, and NEVER omit or reorder an index — the caller re-aligns your response by "entryIndex", not by array position, so a missing or duplicated index silently drops that entry's real classification. rIa Respond with ONLY a JSON array, no markdown, no explanation: [{"entryIndex": , "speakerRole": "DOCTOR" | "PATIENT" | "SMALL_TALK", "mappings": [{"questionId": "", "answer": "", "confidence": }, ...], "customQuestion": "", "continuesFromPrevious": }, ...] Return "mappings": [] when nothing maps, "customQuestion": null when it does not apply, and "continuesFromPrevious": false on every entry unless instructed otherwise above. SMALL_TALKF speakerRolemappingscustomQuestioncontinuesFromPreviousitem question_idsct|tsttS|jd}|dvrd}|jd}t|tsg}g}|D]}t|ts|jd}||vr*|jd}t|t r|j s\ t|jdxsd}|j|||d |jd } |d k7s t| t r| j sd} |jd d u} ||| | dS#ttf$rd}Y|wxYw)Nr)DOCTORPATIENTrrr questionIdrH confidenceg)rrHrrrrTr) isinstancedict_DEFAULT_CLASSIFICATIONr-liststrr!float TypeError ValueErrorrJ) rr speaker_role raw_mappingsrmappingrorHrcustom_questioncontinues_from_previouss r_parse_one_classificationrs] dD !+,,88M*L>># 88J'L lD ) H a'4( kk,/ l * X&&#&flln  w{{<8?C@J {fT^_` ahh/0Oxz/3'GOdOdOf"hh'>?4G$)!8  :& J s;EEEprevious_utterancec2|syd|dd|ddtS)NrDu POSSIBLY INCOMPLETE PREVIOUS UTTERANCE (unresolved — no confident answer matched yet, likely because the speaker paused mid-sentence before finishing): r: r r)_CONTINUATION_INSTRUCTION)rs r_build_previous_utterance_blockrs:  FFXYfFgEhhj f % &b $ % 'rentriesrecent_utterancesc K|sgSdt|}|r/|Dcgc]}|dd|d}}ddj|z}nd}|t|z }djdt|D}|d |d |} t t | d {} t jd d | j} t jdd | } tj| } t!| t"s3tjd| |D cgc]} ttc} S|Dchc]}|d }}i}| D]B}t!|ts t%|j'd}||vr4t-||||<Dt/|t/|k7r*tjdt/|t/|| t1t/|Dcgc]!}|j'|tt#c}Scc}w7z#tj$r<tjd| |D cgc]} ttncc} wc} cYSwxYwcc} wcc}w#t(t*f$rY;wxYwcc}ww)ubClassifies every diarized entry from one audio segment in a single LLM call — one result per entry, in order. entries: [{"speakerId": ..., "text": ...}, ...]. previous_utterance, when given, is the most recent still-unmapped utterance — a candidate for the FIRST entry to continue rather than a fresh turn (see _CONTINUATION_INSTRUCTION).Standard Questions Bank: rrr ,Recent conversation context (oldest first): rCRecent conversation context: (none — this is the first utterance)c3DK|]\}}d|d|dd|dyw)zEntry z (speaker speakerIdz): r Nr )r res rrz.classify_and_map_utterances..s5AEA&:a n-S6 <s rIz* Diarized entries to classify, in order: NrrDrzEclassify_and_map_utterances: could not parse LLM response as JSON: %rzBclassify_and_map_utterances: LLM response was not a JSON array: %rr entryIndexu`classify_and_map_utterances: entryIndex coverage mismatch — %d entries, %d indexed results: %r)rrrrrk_CONVERSATION_CLASSIFY_PROMPTrryr!rrrrvwarningrrrrintr-rrrlenrange)rrrrrNu context_lines context_block entries_blockrVrr0r_rrby_indexr entry_indexrs rclassify_and_map_utterancesrs  23H3S2TUOEVWAm,-R& {;W WG$))TaJbb ^ 45GHHMIIIRSZI[M  4 /44A? D ;\J JCff("ciik:GffZW-G@G$ fd #[]`a7>?!,-?? &//AdG/L/ "H N$%  dhh|45K ( "  9$ M N 8}G $ n L#h- EJ#g,DW XqHLLD!89 : XXgX K   @^`cd7>?!,-???@ @ 0:&    YsJH A%JHAJH+*JI#+J2 I(>JI-3A-J &J J-I >I I JI  J-J<J?JJu_You are a clinical conversation-tagging assistant. You will be shown ONE continuous transcript from a single audio segment of a live, free-flowing conversation during a medical intake. Unlike diarized audio, this transcript has NOT been split by speaker — it is plain text that may contain a doctor's question and the patient's answer (or several such exchanges) run together with no separator, because they were spoken with little pause between them. You are also given the most recent prior classified utterances for context, and a bank of standard intake questions the doctor is expected to cover. YOUR JOB — in order: 1. SPLIT the transcript into consecutive speaker turns, in the order they occur. Cut at the natural boundary between a question and its answer, or between one exchange and the next — for example "What's your problem? I am suffering from fever. When did it start? Two days ago." splits into FOUR turns: "What's your problem?" / "I am suffering from fever." / "When did it start?" / "Two days ago." Keep each turn's wording as close to verbatim as possible — do not paraphrase, only trim leading/trailing filler if needed. If the transcript is genuinely just one turn with nothing to split, return a single segment. 2. For EACH resulting turn, decide the speaker: DOCTOR (asking a question or giving an instruction), PATIENT (the patient or anyone accompanying them describing symptoms or answering), or SMALL_TALK (greetings, chit-chat, or speech unrelated to the intake). Treat any non-doctor speaker — including a family member or companion answering on the patient's behalf — as PATIENT context. 3. For EACH turn, extract EVERY standard question it answers (a turn can answer more than one at once). Only include a mapping when reasonably confident. For each mapped question, extract the direct answer to THAT question only, stripped of filler words. 4. If a turn's speaker is DOCTOR, decide whether it is a genuinely new clinical question NOT covered by the standard bank (a paraphrase of a standard question does not count). If so, return it cleaned of filler words as 'customQuestion', otherwise null. uu OUTPUT: respond with ONLY a JSON array, no markdown, no explanation. Number your turns starting at 0, in the order they occur, via "segmentIndex" — one object per turn: [{"segmentIndex": , "speakerRole": "DOCTOR" | "PATIENT" | "SMALL_TALK", "text": "", "mappings": [{"questionId": "", "answer": "", "confidence": <0.0-1.0>}, ...], "customQuestion": "", "continuesFromPrevious": }, ...] Return "mappings": [] when nothing maps, "customQuestion": null when it does not apply, and "continuesFromPrevious": false on every turn unless instructed otherwise above. transcriptcLK|jsgSdt|}|r/|Dcgc]}|dd|d}}ddj|z}nd}|t|z }|d|d |}t t |d {} t jd d | j} t jd d | } d|gd ddg} tj| } t| tr| stjd| | S|D chc]} | d }} g}| D]}t|ts t!|j#d}|j#d}t|t(r|jsat+||}|j-|i|d|jif|stjd| | S|j/dtt1t3|}|Dcgc]\}}| c}}|k7rtjd| |Dcgc]\}}| c}}Scc}w7#tj$rtjd| | cYSwxYwcc} w#t$t&f$rYswxYwcc}}wcc}}ww)uText-based alternative to audio diarization — takes ONE flat, non-diarized transcript and asks the LLM to both split it into speaker turns AND classify/map each turn, in a single call. Used when Sarvam-side diarization is disabled, so the transcript display and question mapping can still reflect real doctor/patient turn-taking instead of forcing one role onto the whole transcript. previous_utterance, when given, is the most recent still- unmapped utterance — a candidate for the FIRST turn to continue rather than a fresh one. Returns a list of {"speakerRole", "text", "mappings", "customQuestion", "continuesFromPrevious"} in chronological order — same shape classify_and_map_utterances produces, so the caller can treat both paths identically.rrrr rrrrIz$ Transcript to split and classify: NrrDrrF)rr rrrzCsegment_and_map_utterance: could not parse LLM response as JSON: %rzJsegment_and_map_utterance: LLM response was not a non-empty JSON array: %rr segmentIndexzEsegment_and_map_utterance: no valid segments parsed from response: %rc |dS)Nrr )pairs rz+segment_and_map_utterance..ws $q'r)keyuksegment_and_map_utterance: segmentIndex sequence was not contiguous from 0 — using array order anyway: %r)r!rrrrk_CONVERSATION_SEGMENT_PROMPTrryrrrrvrrrrrr-rrrrrJsortrr)rrrrrNrrrrVrr0fallbackrrrindexedr segment_indexr parsed_classificationexpectedrrsegments rsegment_and_map_utterancer4s     23H3S2TUOEVWAm,-R& {;W WG$))TaJbb ^ 45GHHM%&d=/9_`j_klL:LI ICff("ciik:GffZW-G ,jbdhDIJKHG$ fd #6cehi &//AdG/L/&(G Y$%   89Mxx$$DJJL  9$ M 'V*?'V'VWX Y ^`cd LL)L*E#g,'(Hda(* y  '. . 7G ..mX J   \^ab0:&    /s'J$IAJ$I A J$ I !.J$ I=J$6JB6J$ J"J$4 J J$ +I:7J$9I::J$JJ$JJ$rFraw_transcriptcK tdd|d|dd{}|jS7#t$r|cYSwxYww)Nz#You are a medical intake assistant.zQuestion asked to patient: "z" Transcript: "u" Extract ONLY the patient's direct answer to this medical question. Remove background conversation, side talk, and unrelated speech. Preserve all medical details in the answer — including symptoms, duration, timing, severity, and any descriptions the patient gave. If there is no valid patient response at all, return empty string. Return only the extracted response, nothing else.)rkr! Exception)rFrresults rextract_patient_answerrsd $ 1*8*5*+,@ A   ||~   s,A1/1A1 ?A?A)NENGLISH)N)r).__doc__rloggingrrdapp.core.configr getLogger__name__rvrx_SKIP_ANSWEREDrKrLrrrrcompiler IGNORECASE MULTILINErr)tupler3rTrkrqr|rrrrr_DOCTOR_PRIORITY_INSTRUCTIONrrobjectsetrrrrrrr rrrsK  $   8 $)N,OZ.G/G>`T$Z`C`rzz89 2::2BMMBLL4PQ ; ;d ;S ; S T$Z E#PT*VY/DZ ()- %%%%Dz % #Y% %  % %P 7s 7# 7# 7$)- IIIDzI#Y% I  I  I0)-|||| | Dz | #Y% || |@)- &8&8&8&8 &8 #Y% &8  &8V A A$7s7c77(7373737@W)fT!!#$%WW1Wj+7BZ^y~'F'#c('t'Tt '+ CY $ZCYDzCYDzCYt CY  $Z CYN<0!!#$%V3 VR'+ K/K/DzK/DzK/t K/  $Z K/\3r