a qjjb @sbdZddlZddlZddlmZddlmZddlmZddl m Z e Z e dejZe dejZe d ejZd Zhd Zd d ddddgddddgddddgddddgddddgddddddZd dddddidd id!d"d#gd$ddZd d%d&dddidd id'd(d#gd$ddZd)d*Zd+d,Zd-d.Zd/d0ZGd1d2d2ZdS)3uReal agentic loop for the Complaint Guidance Agent — no demo script, no skip flags. Design carried over deliberately from the earlier build (see project memory): the harness decides what step/stage a turn is on from facts it tracks itself, and tells the model that — asking the model to self-report its own stage caused it to drift in the original build. What's new here is that every turn is a live gateway call: fact extraction, response phrasing, and three independent verification passes (citation existence, grounding, relevance), all against the real procedures in knowledge_base/ (see that directory's README for what's genuinely esure-sourced vs. deliberately left out). N)ValidationError) EvidenceStore) GatewayClient) TurnResponsea,\b(reduced income|short time at work|earning less|job loss|lost (my|his|her) job|bereavement|illness|disabilit|caring responsibilit|domestic abuse|financial abuse|can'?t afford|cannot afford|struggling to pay|low confidence|can'?t understand|difficulty understanding|limited literacy|language need)\bzq\b(yes|yeah|yep|agrees?|agreed|accepts?|accepted|happy for|happy to|ok(ay)? with|consents?|confirms?|confirmed)\bz=\b(not|no\b|never|won'?t|don'?t|refus|declin|reject(?!ion))\b ESCALATE_TL>zPROC-EXPLAIN-CANCELLATION#S1zPROC-MANUAL-PAYMENT#S1zPROC-REPRESENT-DD#S1function update_factszRecord which facts the representative has confirmed so far, from their latest note. Only set a field if this note actually confirms it; leave others null.objectintegernullzQHow many Direct Debit collections have been rejected this policy year, if stated.)type descriptionZbooleanzTrue if account notes confirm a previous advisor incorrectly told the customer the DD would be reapplied despite multiple missed payments.z]True if the customer has disclosed a vulnerability indicator (reduced income, illness, etc.).zcTrue/false once the customer has answered whether they consent to the vulnerability being recorded.zMTrue/false once the customer has accepted or rejected the offered resolution.)missed_dd_countincorrect_info_givenvulnerability_disclosedvulnerability_consentcustomer_accepted)r properties)namer parameters)r rgrounding_verdictz_Judge whether every factual claim in the response is supported by the provided source material.r string) supportedreasonrr)r rrequiredrelevance_verdictzPJudge whether the response actually addresses what the representative just said.) responsiverrcCs*|dgD]}|d|kr |Sq dS)NZstepsstep_idget) procedurerstepr"(/home/ec2-user/esure-agent/agent_core.py _step_lookupts  r$c Cs|dD]}|}td|}|s,dS|\}}}||}|durRdSddd|}|durvt|}|dkr||ksdS|dkr||ksdS|d kr ||ks dSq dS) zTiny evaluator for the branch grammar actually used in esure-procedures.json: `FIELD OP VALUE [AND FIELD OP VALUE ...]`, OP in (>=, <, ==). z AND z(\w+)\s*(>=|<|==)\s*(\w+)FNT)truefalsez>=r;cSsg|] }d|iqSZ referencer".0Zaidr"r"r# z&ReasoningAgent.run..rEheadline next_movessuggested_wording instructionZstage citationsZ case_noteanswer verificationr/turn_response_json)dict_VULN_INDICATORSsearch setdefaultrHrG_extract_factsr _AFFIRMATIVE_NEGATION_WORD_escalated_result_instruction_for _context_text_phrase_response_verifyboolrXrYrZ _case_noteZmodel_dump_jsonlen)rKquestionhistorycarriedr/positionr[ evidence_ids context_text turn_responseZgrounding_contextr_rQr^ response_jsonr"r"r#runsh             &   zReasoningAgent.runcCs6d|gd|dgdddgddgd|t|dddS)Nu/Hold the customer — this needs a team leader.r:rWTrQr\ grounding relevance violationsrNr])rp)rKrr/rrr"r"r#ris  z ReasoningAgent._escalated_resultcCs|d}|dkr4t|jd|d}|ddgfS|j|d}t||d}g}|ddkrv|d rv|d|d|fS) Nrr<zPROC-VULNERABILITY-FLOWrNr[zVULNERABLE-CUSTOMER-PROCESSrPrBr;)r$rGr rappend)rKrtr/kindr!r rur"r"r#rj.s zReasoningAgent._instruction_forcCs>g}|D]*}|j|}|r|d|d|qd|S)Nz###  rO)rGZ document_textr~join)rKruparts article_idtextr"r"r#rk:s  zReasoningAgent._context_textc Csdddg}|D]}|d|dddq|d|d|jj|tgdd}t|}|jpbgD]`}|jjd krvqdzt |jj } Wnt j yi} Yn0| D]\} } | dur| || <qqd|S) Nsystema You extract structured facts from a customer-service representative's notes during a missed-Direct-Debit billing complaint call. Only record what this specific note actually states. You MUST call the update_facts tool on every turn, even if nothing new was confirmed.ZrolecontentusernoterMautoZtoolsZ tool_choicer)r~rrJcomplete _FACTS_TOOLrb tool_callsrrjsonloads argumentsJSONDecodeErroritems) rKrqrrr/messagesr`messageZ new_factscallparsedkeyr3r"r"r#rfBs&   zReasoningAgent._extract_factsc Csd|pdd|d|r$d|dnd}d|dg}|D]:}|d |d dd|d r:|d |d dq:|d |d|jj|tgd d} | jpgD]>} | jjdkrqzt | jj WSt yYdS0qdS)uStructured, not free prose — see schemas.TurnResponse for why. Forced tool_choice isn't supported by this model group (see _extract_facts), so this is tool_choice="auto" plus an explicit instruction, same pattern used everywhere else in this file. u\You are a complaint-guidance assistant helping an esure representative handle a live call. You MUST call submit_turn_response with your answer — never reply in plain text. Keep headline to one sentence, next_moves to short imperative bullets, and base everything ONLY on this instruction and source material — never invent a policy detail, number, or process step that isn't in them. This includes UI/system details (e.g. 'dropdown list', specific screens), escalation routes, verification rituals (e.g. 'read the number back'), or any extra precaution/check step (e.g. 'confirm the resolution is still valid before closing') that isn't explicitly stated — if the instruction is terse, keep your answer just as terse rather than filling in plausible-sounding procedure around it, even a well-intentioned one. Only cite ids from this list, never another: z(none provided)z. Instruction for this turn: rOz)Source material you may cite or draw on: rMrrrrrxZ assistantrrZsubmit_turn_responseN) r~rrJr_TURN_RESPONSE_TOOLrrrrZmodel_validate_jsonrr) rKrqrrr[rvrurrr`rrr"r"r#rl_s2        zReasoningAgent._phrase_responsec Csg}|durdgdddgdS|jdd|j|jrBd|jnd}t|j}|D]} | |vr^|d| q^|j D]} | |vr|d| dqdd d |d |igD} d} |r| t d d d| d|d|} | r| ds|d| r| ddnd| t ddd| d|d|} | rP| dsp|d| rh| ddnd| dd|D| | |dS)NFz0model did not return a valid structured responserzrrMz&citation references unknown document: z model cited z!, which wasn't provided this turncss |]}d|ddVqdS)z- rrMNr)rTtr"r"r# rVz)ReasoningAgent._verify..rruDoes every POLICY/PROCESS claim in this response hold up against the source material? The response may also restate facts the representative already reported in the conversation (the customer's own circumstances, what they said, what was checked) — that is NOT a violation, even though those facts aren't in the source material. Only flag a claim about policy, process, or company procedure that is NOT supported by the source material. A paraphrase is fine; an invented policy/process detail is not.zUCONVERSATION SO FAR (facts already confirmed, not to be checked against the source): z> SOURCE MATERIAL (check policy/process claims against this): z RESPONSE: rzgrounding check failed: rzno verdict returnedruJDoes this response follow sensibly from what the representative just said, GIVEN the conversation so far? A short confirmation like "ready" or "yes" is responsive if the response correctly proceeds with the next step implied by the conversation — it does not need to restate or re-explain context the representative already has.zCONVERSATION SO FAR: z# REPRESENTATIVE'S LATEST MESSAGE: rzrelevance check failed: cSsg|] }d|iqSrRr"rSr"r"r#rUrVz*ReasoningAgent._verify..)rXrrYrZsetrGZ documentskeysr~r\_judge_GROUNDING_TOOLr_RELEVANCE_TOOL) rKrqrwrvrurrr}Z response_textZ valid_idsrZ conversationr{r|r"r"r#rmsr      zReasoningAgent._verifyc Cs|d|d}|jjd|dd|dg|gdd}|jp}|jj|kr>zt|jjWStjyzYdS0q>dS)Nz You MUST call the z tool with your verdict.rrrrr) rJrrrrrrrr)rKZtoolZ tool_nameZ system_promptZ user_contentrrr"r"r#rs  zReasoningAgent._judgecCs`d|ddd|dr dndd|d rDd |d d nd d|drZdndS)NzXCustomer complaint: Customer complained after receiving a cancellation notice following rZmultiplez( missed Direct Debits. Account review: rzNPrevious advisor notes confirmed incorrect DD-reapplication advice was given. z)No incorrect prior advice was confirmed. z Vulnerability: rz,Vulnerability disclosed; consent to record: rz. zNo vulnerability disclosed. z Customer response: rz"Customer accepted the resolution. z%Customer response not yet confirmed. r)rKr/r"r"r#ros&   zReasoningAgent._case_note)NN)NN) __name__ __module__ __qualname__rLryrirjrkrfrlrmrror"r"r"r#rIs  G (UrI)__doc__rr*ZpydanticrZevidence_storerZ model_clientrZschemasrZ tool_schemarcompile IGNORECASErcrgrhrrFrrrr$r4r7rHrIr"r"r"r#s     %5