KCj* dZddlZddlZddlmZddlmZeZeeZ dZ de de de d e d e f d Zde de de d e d e f d ZGd dZeZy)u  LLM reasoning layer. Production concerns specific to calling an external LLM API: 1. Timeouts — never let one slow LLM call hold a worker indefinitely. 2. Retries — transient network/5xx errors should retry with backoff, but only a bounded number of times (LLM_MAX_RETRIES), not forever. 3. Graceful degradation — if the LLM is unreachable after retries, the user should still get the raw model numbers with a template-based explanation, not a 502 for the entire request. The ML predictions are the product; the LLM is a presentation layer on top of them. 4. No hallucination — the system prompt explicitly forbids inventing numbers, and the user prompt only contains values that came from the aggregator, never free text the model could "fill in." N) get_settings) get_loggeruYou are an agricultural decision assistant helping Indian farmers and agricultural officers understand machine learning model output about crop suitability. Strict rules you must follow: - Use ONLY the numbers and labels provided to you. Never invent or estimate a number that was not given. - Do not override or contradict the model outputs. Your job is to explain them, not to second-guess them. - If a model output is marked "degraded" or has null values, say so plainly rather than guessing a substitute number. - If conflicts are listed, mention them honestly — do not smooth over genuine uncertainty. - Use simple, direct language suitable for someone without a technical background. Avoid jargon like "anomaly score" without explaining what it means. - Structure your response in exactly six sections, in this order: 1. Crop suitability summary 2. Climate assessment 3. Expected yield 4. Market insight 5. Risk explanation 6. Alternative crop suggestions - Keep the whole response under 350 words. districtcropmonth model_outputsreturncd|d|d|d|dd|dd|d d |d d |d d|dd|dd|jdgdS)NzCrop: z District: z Month: z Model outputs: - Climate: climatez - Feasibility: feasibilityz - Yield: yield_predictionz - Trend: trendz - Market: marketz - Alternative crops: recommendationz - Conflicts detected: conflictsz$ - Models that had issues this run: degraded_modelszL Write the six-section explanation now, following the system rules exactly.)get)rrrrs EC:\Crop_Prediction\Backend\crop-ai-system\app\services\llm_service.py_build_user_promptr5sdV * w  ) $ %&m,-. * + ,-  !"  " #$#$4567$[123$$1$5$56G$L#MNK NNc|d}|d}|d}|d}|djdg}|rdjd|Dnd } d |d |d |d |jddd|jdddd|jdd|jdd|jdd|jdd|jdddj|jd gxsd!d"| d#S)$u Used only if the LLM is unreachable after all retries. A plain template ensures the user still receives the actual model results — degraded presentation, not a failed request. r r r rr alternativesz, c3&K|] }|d yw)rN).0as r z(_fallback_explanation..Rs2!F)2sznone availablezlAI explanation service is temporarily unavailable, showing raw model results. 1. Crop suitability summary: z in z for month z is classified as 'labelunknownz' (confidence confidencerz.0%z,). 2. Climate assessment: forecast rainfall forecast_rainfall_mmz mm, anomaly level: anomaly_labelz. 3. Expected yield: expected_yield_ton_haz3 tonnes/hectare. 4. Market insight: price forecast price_forecast_per_quintalz per quintal, demand level: demand_levelz. 5. Risk explanation: z; rzno conflicts detectedz#. 6. Alternative crop suggestions: .)rjoin) rrrrfeasyldr ralts alt_namess r_fallback_explanationr,GsL  'D * +CI&G 8 $F ) * . .~r BD6: 2T22@PI ((,vT(;ugN((7I678xx a05644;KK@V4W3XY!++o678!gg&=>?@--3ZZ8T-U,VW&&,jj&@%AB $ -*;*;K*L M hQhij,,5;a 9 rc 8eZdZd dZd dZdededededef d Zy) LLMServicer Ncd|_y)N)_clientselfs r__init__zLLMService.__init__ds 37 rctjstjdd|_yt j tjtj|_y)NukANTHROPIC_API_KEY is not set — LLM explanations will use the fallback template until a key is configured.)api_keytimeout)settingsANTHROPIC_API_KEYloggerwarningr0 anthropic AnthropicLLM_TIMEOUT_SECONDSr1s rinitzLLMService.initgsM)) NNC  DL  **..00  rrrrrc |jt||||St||||}d}tdtj dzD]} |jj jtjtjtd|dg}|jD cgc]} | jdk(s| j } } dj| jcSt&j5d|t||||Scc} w#t j"$rM} | }t%d|zd} t&j)d | d |d t+j,| Yd} ~ ,d} ~ wt j.$rM} | }t%d|zd} t&j)d | d |d t+j,| Yd} ~ d} ~ wt j0$r} | }d | j2cxkrdkrNnnKt%d|zd} t&j)d| j2d| dt+j,| n,t&j5d| j2d| Yd} ~ Yd} ~ 7d} ~ wwxYw)Nuser)rolecontent)model max_tokenssystemmessagestext zLLM rate limited, retrying in z s (attempt )z"LLM connection error, retrying in iiXzLLM server error z, retrying in szLLM client error z, not retrying: z;LLM call failed after retries, using fallback explanation: )r0r,rranger7LLM_MAX_RETRIESrHcreate LLM_MODELLLM_MAX_TOKENS SYSTEM_PROMPTrDtyperIr'stripr;RateLimitErrorminr9r:timesleepAPIConnectionErrorAPIStatusError status_codeerror) r2rrrr user_prompt last_errorattemptresponseb text_blocksexcwaits rget_explanationzLLMService.get_explanationtsG << (4 N N(4 N '+ Q 8 81 <=! G <<0077",,'66('-+FG 8 08/?/?T!166VCSqvvT Tyy-3355! F  RS]R^_`$XtUMJJ9U++ ! 1<,!?v[QXPYYZ[\ 4  // ! 1<,!CD6U\T]]^_` 4  ++  #///C/qG|R0DNN%6s6G~VZU[[\#]^JJt$LL#4S__4EEUVYUZ![\ % sL AD'D<D DDJ,AE44J AGJ(B I??J)r N) __name__ __module__ __qualname__r3r>strintdictrfrrrr.r.cs<8  +K+K3+Ks+KSW+K\_+Krr.)__doc__rXr;app.core.configrapp.core.logging_configrr7rgr9rSrjrkrlrr,r. llm_servicerrrrqs  (. > H  4NNCNNDNUXN$Cs3tX[8<K<K@l r