AiXddlZddlZddlmZmZddlZddlmZddl m Z m Z m Z m Z mZmZmZmZmZmZddlmZee Zdedejfd Zd ed eedefd Zd edeeefdefdZdededeefdZdefdZedkreZ e!de dSdS)N)ListDict)OpenAI) OPENAI_API_KEY EMBED_MODEL LLM_MODEL INDEX_DIR ARTIFACTS_DIR CATEGORIESTOP_K_PER_CATEGORYMAX_FINDINGS_PER_CATEGORYRUN_ID INPUT_DIR) FaissStore)api_keyqueryreturnctjt|g}t j|jdjgdS)z.Convert a text query into an embedding vector.)modelinputrfloat32)dtype)client embeddingscreaternparraydata embedding)rresps DC:\Users\Terasoftware\OneDrive\Desktop\Graph Rag\rag_suite\rag\qa.py embed_queryr"sJ   # #g $  D 8TYq\+,I > > >>questionsearch_resultsc h|sdSg}|D]x}|di}|dd}|dd}|dd}|d |d |d |yd |d d}d|d|d dd} tjjtd| dgdd} | jdj j S#t$r} dt| cYd } ~ Sd } ~ wwxYw)z\ Generate an LLM answer based on retrieved chunks. Used by main.py ask command. z/No relevant information found in the documents.metadatadoc_nameUnknown page_number?text[z, Page z] z NzUYou are an AI assistant analyzing annual reports and financial documents. Question: z Context from documents: i.a Instructions: - Provide a clear, factual answer based ONLY on the context above - Cite sources using [Document Name, Page X] format after each claim - If the information is not in the context, say "Information not available in the provided documents" - Be concise but complete Answer:userrolecontentg?iX)rmessages temperature max_tokensrzError generating answer: )getappendjoinrchat completionsrrchoicesmessager3strip Exceptionstr) r$r% context_partsrmetar(pager,contextpromptresponsees r!answer_question_with_llmrI(s A@@M CCuuZ$$88J 22xx s++xx##AAA$AA4AABBBBkk-,--G      %    F 4;*11%&99: 2   "*288::: 44433q66333333334s1AD D1D,&D1,D1evidence chunk_mapcv||d}|sdS|dd}|sdS||dvrdS|d|dkrdS|d|dkrdSd S) z Validate that evidence actually exists in the retrieved chunks. Checks: chunk_id exists, quote is in text, doc_name matches, page matches. chunk_idFquoter-r,r(rDr*T)r7lower)rJrKchunkrNs r!validate_evidencerQYs MM(,,z22 3 3E u LL" % % + + - -E u E&M'')))u||J5#44u||Fu]33u 4r#categoryctjt}t|}||t }g}i|D]P}|d}|d|d|d|d|dd}||||d<Qtj td | d d  d d d } tj tdt| dd5} tj|| ddddn #1swxYwYd|dtj|dd|d} t$jjt,d| dgd} tj| jdjj} n:#tjt8f$r!}t;d|d|gcYd}~Sd}~wwxYwg}| d gdt>D]@} fd!| d"gD}|r|| d"<|| A|S)#z For a given category (e.g., 'Risks', 'Management'), retrieve relevant chunks and use LLM to extract structured findings with evidence. )kr'rMr(r* source_urlr,)rMr(r*rUr,retrieval_bundle_ _/z.jsonTexist_okwutf-8encoding)indentNzD You are a pharma-grade annual report extraction engine. Category: u RULES: - Return VALID JSON ONLY - Use ONLY the provided chunks - If unsupported, OMIT the finding - Each finding must include evidence with doc_name, page, chunk_id, quote (≤25 words) CHUNKS: z" OUTPUT SCHEMA: { "category": "aC", "findings": [ { "finding": "Short description", "severity": "High|Medium|Low", "confidence": "High|Medium|Low", "evidence": [ { "doc_name": "...", "page": 123, "chunk_id": "...", "quote": "exact quote from text" } ] } ] } r0r1r)rr4r5u%⚠️ Error extracting findings for z: findingsc4g|]}t||S)rQ).0rHrKs r! z1extract_findings_per_category..s9    I..    r#rJ) rloadr r"searchr r8ospathr9r rOreplacemakedirsopenjsondumpdumpsrr:r;rrloadsr<r=r3JSONDecodeErrorr?printr7r )rRrstoreqvechitsbundlehrCentry bundle_pathfrFr resultrHrbvalid_evidencerKs @r!extract_findings_per_categoryr~us OI & &E u  D << 2< 3 3D FI  , ,}Z(Z( .|,L     e&+ $z"##',,WHNN,,44S#>>FFsCPPWWWKK ---- k3 1 1 1'Q &!A&&&&'''''''''''''''! !!F1!! !!!!FF {&--%&99:.   DLO3;<<  ) , EhEE!EEFFF H ZZ B ' '(B)B(B C    uuZ,,     *AjM OOA    Os1-EEE=AGH-H H HcVg}d}tdtjD]u\}}td|dt||}|D]*}d|d|d<||d<|dz }||+td t |d v|D]e}d |d D}t ||d <dddd|ddd}|d|d zz|d<f|ddd|ddD}ttt |||d} tj tdtj td} t| dd5}t!j| |dd !dddn #1swxYwYtd"t |td#d$|td%| d&| S)'z Generate the master findings.json file by: 1. Extracting findings for each category 2. Scoring and ranking findings 3. Identifying top 10 critical findings u* 🔍 Extracting findings per category... z Processing: z...zF-03didrRu → Found z findingsch|] }|d S)r(rd)rerHs r! z)generate_findings_json..s888Q1Z=888r#rJrecurrence_countr`)HighMediumLowseverityrscorec|dS)Nrrd)xs r!z(generate_findings_json..s AgJr#T)keyreversecg|] }|d S)rrd)rer{s r!rfz*generate_findings_json..s555a!D'555r#N )run_id input_foldertotal_findingsrb top_10_idsrZz findings.jsonr\r]r^F)ra ensure_asciiu ✅ Total findings: u✅ Top 10 critical findings: z, u✅ findings.json saved at:  )rsr itemsr~r8lenr7sortrrrirlr rjr9rmrnro) all_findingsfidrRrrbr{doc_set sev_scorer findings_dataout_paths r!generate_findings_jsonrsL C 8999%+--99% ,x,,,---05AA # #A$3nnnAdG$AjM 1HC    " " " " 7s8}}7778888==88!J-888 #G  !A66::155U;S;SUVWW !a(:&;";<' ..===55<#4555J!l++  MK ----w||M?;;H hg . . .B! -15AAAABBBBBBBBBBBBBBB 63|#4#4 6 6777 B499Z+@+@ B BCCC 5 5 5 5666 Os"GG G __main__u✅ findings.json generated at )"rnritypingrrnumpyropenair rag.configrrrr r r r r rrrag.vectorstorerrr@ndarrayr"rIboolrQr~r__name__rjrsrdr#r!rs                         '&&&&&  ' ' ' ?s?rz????+4s+4DJ+43+4+4+4+4bc4iT8ZCZZT ZZZZ@22222j z4 ! ! # #D E 2D 2 23333344r#