po/Dokumenten-Check
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Kosten Endpunkt

Browse Source
This commit is contained in:
Wolf G. Beckmann
2026-06-01 14:21:42 +02:00
parent fb185c1a71
commit f3ae6859bf
21 changed files with 632 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
.claude/settings.local.json
View File

@@ -33,7 +33,8 @@
"Bash(NLS_LANG=GERMAN_GERMANY.AL32UTF8 sql *)",
"Bash(KUBECONFIG=/mnt/c/source/Frigosped/Dokumenten-Check/env/frigo-dev.yaml kubectl logs *)",
"Bash(KUBECONFIG=/mnt/c/source/Frigosped/Dokumenten-Check/env/frigo-dev.yaml kubectl get pods -n ai-env -l app=dc-backend)",
"Bash(chmod +x *)"
"Bash(chmod +x *)",
"Bash(JAVA_HOME=/usr/lib/jvm/java-21-openjdk-amd64 mvn compile -q)"
]
}
}

143
Scripts/KI-Kostenverfolgung.sql Normal file
View File

@@ -0,0 +1,143 @@
-- =============================================================================
-- KI-Kostenverfolgung für Dokumenten-Check
-- Tabelle: dc_ai_cost_log
-- Package: DC_COSTS_PKG
-- =============================================================================
-- Deployment:
-- sqlplus user/pass@db @"Scripts/KI-Kostenverfolgung.sql"
--
-- Skript ist idempotent (DROP IF EXISTS vor jedem CREATE).
-- =============================================================================
-- =============================================================================
-- Tabelle: dc_ai_cost_log
-- =============================================================================
BEGIN
EXECUTE IMMEDIATE 'DROP TABLE dc_ai_cost_log CASCADE CONSTRAINTS';
EXCEPTION WHEN OTHERS THEN NULL;
END;
/
CREATE TABLE dc_ai_cost_log (
id NUMBER GENERATED BY DEFAULT AS IDENTITY PRIMARY KEY,
called_at TIMESTAMP DEFAULT SYSTIMESTAMP NOT NULL,
provider VARCHAR2(30) NOT NULL, -- mistral | ollama
model_name VARCHAR2(100) NOT NULL,
operation VARCHAR2(30) NOT NULL, -- TRANSLATE | EVALUATE | OCR | CATALOG
project_id NUMBER,
catalog_id NUMBER,
prompt_tokens NUMBER DEFAULT 0 NOT NULL,
completion_tokens NUMBER DEFAULT 0 NOT NULL,
total_tokens NUMBER DEFAULT 0 NOT NULL,
cost_eur NUMBER(14, 8) DEFAULT 0 NOT NULL,
created_at TIMESTAMP DEFAULT SYSTIMESTAMP NOT NULL,
created_by VARCHAR2(200) DEFAULT 'SYSTEM' NOT NULL
);
COMMENT ON TABLE dc_ai_cost_log IS 'Protokoll aller KI-API-Aufrufe mit Token-Verbrauch und Kosten';
COMMENT ON COLUMN dc_ai_cost_log.operation IS 'TRANSLATE, EVALUATE, OCR, CATALOG';
COMMENT ON COLUMN dc_ai_cost_log.cost_eur IS 'Berechnete Kosten in EUR basierend auf konfigurierten Preisen';
CREATE INDEX idx_ai_cost_called_at ON dc_ai_cost_log (called_at);
CREATE INDEX idx_ai_cost_project_id ON dc_ai_cost_log (project_id);
CREATE INDEX idx_ai_cost_catalog_id ON dc_ai_cost_log (catalog_id);
CREATE INDEX idx_ai_cost_ym ON dc_ai_cost_log (TO_CHAR(called_at, 'YYYY-MM'));
-- =============================================================================
-- PL/SQL Package: DC_COSTS_PKG
-- =============================================================================
CREATE OR REPLACE PACKAGE DC_COSTS_PKG AS
-- -------------------------------------------------------------------------
-- Gesamtkosten + Kosten des laufenden Monats
-- -------------------------------------------------------------------------
PROCEDURE get_summary(
p_total_eur OUT NUMBER,
p_total_tokens OUT NUMBER,
p_total_calls OUT NUMBER,
p_month_eur OUT NUMBER,
p_month_tokens OUT NUMBER,
p_month_calls OUT NUMBER
);
-- -------------------------------------------------------------------------
-- Kosten pro Monat + Modell (für APEX-Reports / direkte Abfragen)
-- Gibt einen REF CURSOR zurück (Spalten: month, provider, model_name,
-- operation, call_count, prompt_tokens, completion_tokens, total_tokens, cost_eur)
-- -------------------------------------------------------------------------
FUNCTION get_monthly_detail RETURN SYS_REFCURSOR;
END DC_COSTS_PKG;
/
CREATE OR REPLACE PACKAGE BODY DC_COSTS_PKG AS
PROCEDURE get_summary(
p_total_eur OUT NUMBER,
p_total_tokens OUT NUMBER,
p_total_calls OUT NUMBER,
p_month_eur OUT NUMBER,
p_month_tokens OUT NUMBER,
p_month_calls OUT NUMBER
) IS
BEGIN
-- Gesamtkosten (all time)
SELECT NVL(SUM(cost_eur), 0),
NVL(SUM(total_tokens), 0),
COUNT(*)
INTO p_total_eur, p_total_tokens, p_total_calls
FROM dc_ai_cost_log;
-- Laufender Monat
SELECT NVL(SUM(cost_eur), 0),
NVL(SUM(total_tokens), 0),
COUNT(*)
INTO p_month_eur, p_month_tokens, p_month_calls
FROM dc_ai_cost_log
WHERE called_at >= TRUNC(SYSDATE, 'MM');
EXCEPTION
WHEN NO_DATA_FOUND THEN
p_total_eur := 0; p_total_tokens := 0; p_total_calls := 0;
p_month_eur := 0; p_month_tokens := 0; p_month_calls := 0;
END get_summary;
FUNCTION get_monthly_detail RETURN SYS_REFCURSOR IS
v_cur SYS_REFCURSOR;
BEGIN
OPEN v_cur FOR
SELECT
TO_CHAR(called_at, 'YYYY-MM') AS month,
provider,
model_name,
operation,
COUNT(*) AS call_count,
SUM(prompt_tokens) AS prompt_tokens,
SUM(completion_tokens) AS completion_tokens,
SUM(total_tokens) AS total_tokens,
ROUND(SUM(cost_eur), 6) AS cost_eur
FROM dc_ai_cost_log
WHERE called_at >= ADD_MONTHS(TRUNC(SYSDATE, 'MM'), -11)
GROUP BY TO_CHAR(called_at, 'YYYY-MM'), provider, model_name, operation
ORDER BY month DESC, cost_eur DESC;
RETURN v_cur;
END get_monthly_detail;
END DC_COSTS_PKG;
/
-- =============================================================================
-- Kurztest (optionale Verifikation nach Deployment)
-- =============================================================================
-- DECLARE
-- v_te NUMBER; v_tt NUMBER; v_tc NUMBER;
-- v_me NUMBER; v_mt NUMBER; v_mc NUMBER;
-- BEGIN
-- DC_COSTS_PKG.get_summary(v_te,v_tt,v_tc,v_me,v_mt,v_mc);
-- DBMS_OUTPUT.PUT_LINE('Total EUR: ' || v_te || ', Calls: ' || v_tc);
-- DBMS_OUTPUT.PUT_LINE('Month EUR: ' || v_me || ', Calls: ' || v_mc);
-- END;
-- /

130
Scripts/ORDS REST Services.sql
View File

@@ -1026,6 +1026,136 @@ BEGIN
]'
);
-- ===========================================================================
-- TEMPLATE 17: ai-costs (POST = Eintrag speichern, GET = Auswertung)
-- ===========================================================================
ORDS.DEFINE_TEMPLATE(
p_module_name => 'frigosped.dc',
p_pattern => 'ai-costs',
p_priority => 0,
p_comments => 'KI-Kostenprotokoll: POST speichert Aufruf, GET liefert Auswertung'
);
-- 17a. POST /api/dc/ai-costs
-- Body: {"provider":"mistral","model_name":"mistral-small-latest","operation":"TRANSLATE",
-- "project_id":42,"catalog_id":null,
-- "prompt_tokens":1234,"completion_tokens":567,"total_tokens":1801,"cost_eur":0.000157}
ORDS.DEFINE_HANDLER(
p_module_name => 'frigosped.dc',
p_pattern => 'ai-costs',
p_method => 'POST',
p_source_type => ORDS.SOURCE_TYPE_PLSQL,
p_comments => '201 Created mit neuer cost_log id',
p_source => q'[
DECLARE
v_body CLOB;
v_new_id NUMBER;
BEGIN
v_body := :body_text;
INSERT INTO dc_ai_cost_log (
provider, model_name, operation,
project_id, catalog_id,
prompt_tokens, completion_tokens, total_tokens,
cost_eur
) VALUES (
JSON_VALUE(v_body, '$.provider'),
JSON_VALUE(v_body, '$.model_name'),
JSON_VALUE(v_body, '$.operation'),
TO_NUMBER(JSON_VALUE(v_body, '$.project_id')),
TO_NUMBER(JSON_VALUE(v_body, '$.catalog_id')),
NVL(TO_NUMBER(JSON_VALUE(v_body, '$.prompt_tokens')), 0),
NVL(TO_NUMBER(JSON_VALUE(v_body, '$.completion_tokens')), 0),
NVL(TO_NUMBER(JSON_VALUE(v_body, '$.total_tokens')), 0),
NVL(TO_NUMBER(JSON_VALUE(v_body, '$.cost_eur')), 0)
)
RETURNING id INTO v_new_id;
:status_code := 201;
HTP.P('{"id":' || v_new_id || '}');
EXCEPTION
WHEN OTHERS THEN
ROLLBACK;
:status_code := 500;
HTP.P('{"error":"' || REPLACE(SQLERRM, '"', '''') || '"}');
END;
]'
);
-- 17b. GET /api/dc/ai-costs
-- Liefert: Gesamtkosten, laufender Monat, Monats-Aufschlüsselung (12 Monate)
ORDS.DEFINE_HANDLER(
p_module_name => 'frigosped.dc',
p_pattern => 'ai-costs',
p_method => 'GET',
p_source_type => ORDS.SOURCE_TYPE_PLSQL,
p_comments => 'Kostenauswertung: total, current_month, monatliche Aufschluesselung via DC_COSTS_PKG',
p_source => q'[
DECLARE
v_total_eur NUMBER;
v_total_tokens NUMBER;
v_total_calls NUMBER;
v_month_eur NUMBER;
v_month_tokens NUMBER;
v_month_calls NUMBER;
v_monthly_json CLOB;
BEGIN
DC_COSTS_PKG.get_summary(
v_total_eur, v_total_tokens, v_total_calls,
v_month_eur, v_month_tokens, v_month_calls
);
SELECT JSON_ARRAYAGG(
JSON_OBJECT(
'month' VALUE month,
'cost_eur' VALUE ROUND(cost_eur, 6),
'prompt_tokens' VALUE prompt_tokens,
'completion_tokens' VALUE completion_tokens,
'total_tokens' VALUE total_tokens,
'call_count' VALUE call_count
)
ORDER BY month DESC
RETURNING CLOB
)
INTO v_monthly_json
FROM (
SELECT
TO_CHAR(called_at, 'YYYY-MM') AS month,
SUM(cost_eur) AS cost_eur,
SUM(prompt_tokens) AS prompt_tokens,
SUM(completion_tokens) AS completion_tokens,
SUM(total_tokens) AS total_tokens,
COUNT(*) AS call_count
FROM dc_ai_cost_log
WHERE called_at >= ADD_MONTHS(TRUNC(SYSDATE, 'MM'), -11)
GROUP BY TO_CHAR(called_at, 'YYYY-MM')
);
:status_code := 200;
HTP.P(
'{"total":{'
|| '"cost_eur":' || ROUND(NVL(v_total_eur, 0), 6)
|| ',"total_tokens":'|| NVL(v_total_tokens, 0)
|| ',"call_count":' || NVL(v_total_calls, 0)
|| '},"current_month":{'
|| '"month":"' || TO_CHAR(SYSDATE, 'YYYY-MM') || '"'
|| ',"cost_eur":' || ROUND(NVL(v_month_eur, 0), 6)
|| ',"total_tokens":'|| NVL(v_month_tokens, 0)
|| ',"call_count":' || NVL(v_month_calls, 0)
|| '},"monthly":' || NVL(v_monthly_json, '[]')
|| '}'
);
EXCEPTION
WHEN OTHERS THEN
:status_code := 500;
HTP.P('{"error":"' || REPLACE(SQLERRM, '"', '''') || '"}');
END;
]'
);
COMMIT;
END;
/

5
dc-backend/pom.xml
View File

@@ -73,6 +73,11 @@
<artifactId>langchain4j-ollama</artifactId>
<version>${langchain4j.version}</version>
</dependency>
<dependency>
<groupId>dev.langchain4j</groupId>
<artifactId>langchain4j-mistral-ai</artifactId>
<version>${langchain4j.version}</version>
</dependency>
<!-- ================================================================ -->
<!-- Dokumentenverarbeitung -->

129
dc-backend/src/main/java/de/frigosped/dc/ai/AiModelProducer.java
View File

@@ -1,6 +1,7 @@
package de.frigosped.dc.ai;
import dev.langchain4j.model.chat.ChatLanguageModel;
import dev.langchain4j.model.mistralai.MistralAiChatModel;
import dev.langchain4j.model.ollama.OllamaChatModel;
import jakarta.enterprise.context.ApplicationScoped;
import jakarta.enterprise.inject.Produces;
@@ -10,29 +11,42 @@ import org.eclipse.microprofile.config.inject.ConfigProperty;
import java.time.Duration;
import java.util.Map;
/**
* CDI-Producer für die zwei KI-Modelle.
*
* Beide Modelle sprechen denselben Ollama-kompatiblen Endpunkt an,
* unterscheiden sich aber in Modellname und Timeout.
* Der X-API-KEY-Header wird über customHeaders gesetzt.
*
* Einsatz:
* @Inject @Named("main") ChatLanguageModel mainModel → Auswertung + Übersetzung
* @Inject @Named("ocr") ChatLanguageModel ocrModel → PDF-OCR (Vision)
*/
@ApplicationScoped
public class AiModelProducer {
@ConfigProperty(name = "dc.ai.base-url")
String baseUrl;
@ConfigProperty(name = "dc.ai.provider", defaultValue = "ollama")
String provider;
@ConfigProperty(name = "dc.ai.api-key")
String apiKey;
// --- Ollama ---
@ConfigProperty(name = "dc.ai.ollama.base-url", defaultValue = "")
String ollamaBaseUrl;
@ConfigProperty(name = "dc.ai.main.model")
String mainModelName;
@ConfigProperty(name = "dc.ai.ollama.api-key", defaultValue = "")
String ollamaApiKey;
@ConfigProperty(name = "dc.ai.ollama.main.model", defaultValue = "")
String ollamaMainModel;
@ConfigProperty(name = "dc.ai.ollama.catalog.model", defaultValue = "")
String ollamaCatalogModel;
@ConfigProperty(name = "dc.ai.ollama.ocr.model", defaultValue = "")
String ollamaOcrModel;
// --- Mistral ---
@ConfigProperty(name = "dc.ai.mistral.api-key", defaultValue = "")
String mistralApiKey;
@ConfigProperty(name = "dc.ai.mistral.main.model", defaultValue = "mistral-small-latest")
String mistralMainModel;
@ConfigProperty(name = "dc.ai.mistral.catalog.model", defaultValue = "mistral-small-latest")
String mistralCatalogModel;
@ConfigProperty(name = "dc.ai.mistral.ocr.model", defaultValue = "pixtral-12b-2409")
String mistralOcrModel;
// --- Gemeinsame Parameter ---
@ConfigProperty(name = "dc.ai.main.timeout", defaultValue = "300s")
String mainTimeout;
@@ -42,9 +56,6 @@ public class AiModelProducer {
@ConfigProperty(name = "dc.ai.main.temperature", defaultValue = "0.1")
double mainTemperature;
@ConfigProperty(name = "dc.ai.catalog.model")
String catalogModelName;
@ConfigProperty(name = "dc.ai.catalog.timeout", defaultValue = "1200s")
String catalogTimeout;
@@ -54,9 +65,6 @@ public class AiModelProducer {
@ConfigProperty(name = "dc.ai.catalog.temperature", defaultValue = "0.1")
double catalogTemperature;
@ConfigProperty(name = "dc.ai.ocr.model")
String ocrModelName;
@ConfigProperty(name = "dc.ai.ocr.timeout", defaultValue = "120s")
String ocrTimeout;
@@ -69,68 +77,89 @@ public class AiModelProducer {
@ConfigProperty(name = "dc.ai.log-responses", defaultValue = "false")
boolean logResponses;
/**
* Hauptmodell: Auswertung der Fragen + Übersetzung
* Modell: gpt-oss:20b
*/
@Produces
@ApplicationScoped
@Named("main")
public ChatLanguageModel mainModel() {
return OllamaChatModel.builder()
.baseUrl(baseUrl)
.modelName(mainModelName)
if (isMistral()) {
return MistralAiChatModel.builder()
.apiKey(mistralApiKey)
.modelName(mistralMainModel)
.temperature(mainTemperature)
.timeout(parseDuration(mainTimeout))
.maxRetries(mainMaxRetries)
.customHeaders(Map.of("X-API-KEY", apiKey))
.logRequests(logRequests)
.logResponses(logResponses)
.build();
}
return OllamaChatModel.builder()
.baseUrl(ollamaBaseUrl)
.modelName(ollamaMainModel)
.temperature(mainTemperature)
.timeout(parseDuration(mainTimeout))
.maxRetries(mainMaxRetries)
.customHeaders(Map.of("X-API-KEY", ollamaApiKey))
.logRequests(logRequests)
.logResponses(logResponses)
.build();
}
/**
* Katalog-Modell: Vollständige Katalog-Generierung aus Regelwerk-Dokument.
* Längerer Timeout (1200s) notwendig da komplette Dokumente analysiert werden.
*/
@Produces
@ApplicationScoped
@Named("catalog")
public ChatLanguageModel catalogModel() {
return OllamaChatModel.builder()
.baseUrl(baseUrl)
.modelName(catalogModelName)
if (isMistral()) {
return MistralAiChatModel.builder()
.apiKey(mistralApiKey)
.modelName(mistralCatalogModel)
.temperature(catalogTemperature)
.timeout(parseDuration(catalogTimeout))
.maxRetries(catalogMaxRetries)
.customHeaders(Map.of("X-API-KEY", apiKey))
.logRequests(logRequests)
.logResponses(logResponses)
.build();
}
return OllamaChatModel.builder()
.baseUrl(ollamaBaseUrl)
.modelName(ollamaCatalogModel)
.temperature(catalogTemperature)
.timeout(parseDuration(catalogTimeout))
.maxRetries(catalogMaxRetries)
.customHeaders(Map.of("X-API-KEY", ollamaApiKey))
.logRequests(logRequests)
.logResponses(logResponses)
.build();
}
/**
* OCR-Modell: Bildtext-Extraktion aus PDF-Seiten (Vision-Modus)
* Modell: quen3.5:9b
*/
@Produces
@ApplicationScoped
@Named("ocr")
public ChatLanguageModel ocrModel() {
return OllamaChatModel.builder()
.baseUrl(baseUrl)
.modelName(ocrModelName)
if (isMistral()) {
return MistralAiChatModel.builder()
.apiKey(mistralApiKey)
.modelName(mistralOcrModel)
.timeout(parseDuration(ocrTimeout))
.maxRetries(ocrMaxRetries)
.customHeaders(Map.of("X-API-KEY", apiKey))
.logRequests(logRequests)
.logResponses(logResponses)
.build();
}
return OllamaChatModel.builder()
.baseUrl(ollamaBaseUrl)
.modelName(ollamaOcrModel)
.timeout(parseDuration(ocrTimeout))
.maxRetries(ocrMaxRetries)
.customHeaders(Map.of("X-API-KEY", ollamaApiKey))
.logRequests(logRequests)
.logResponses(logResponses)
.build();
}
/**
* Parst Timeout-Strings wie "300s", "5m", "120s" in Duration.
*/
private boolean isMistral() {
return "mistral".equalsIgnoreCase(provider);
}
private Duration parseDuration(String value) {
if (value.endsWith("s")) {
return Duration.ofSeconds(Long.parseLong(value.replace("s", "")));

12
dc-backend/src/main/java/de/frigosped/dc/client/OrdsClient.java
View File

@@ -179,4 +179,16 @@ public interface OrdsClient {
@DELETE
@Path("/projects/{projectId}/results")
Response deleteResults(@PathParam("projectId") long projectId);
// =========================================================================
// KI-Kostenverfolgung
// =========================================================================
/**
* POST /api/dc/ai-costs
* Speichert einen KI-Aufruf mit Token-Verbrauch und Kosten in dc_ai_cost_log.
*/
@POST
@Path("/ai-costs")
Response logAiCost(AiCostRequest body);
}

41
dc-backend/src/main/java/de/frigosped/dc/model/AiCostRequest.java Normal file
View File

@@ -0,0 +1,41 @@
package de.frigosped.dc.model;
public class AiCostRequest {
private String provider;
private String modelName;
private String operation;
private Long projectId;
private Long catalogId;
private int promptTokens;
private int completionTokens;
private int totalTokens;
private double costEur;
public String getProvider() { return provider; }
public void setProvider(String v) { this.provider = v; }
public String getModelName() { return modelName; }
public void setModelName(String v) { this.modelName = v; }
public String getOperation() { return operation; }
public void setOperation(String v) { this.operation = v; }
public Long getProjectId() { return projectId; }
public void setProjectId(Long v) { this.projectId = v; }
public Long getCatalogId() { return catalogId; }
public void setCatalogId(Long v) { this.catalogId = v; }
public int getPromptTokens() { return promptTokens; }
public void setPromptTokens(int v) { this.promptTokens = v; }
public int getCompletionTokens() { return completionTokens; }
public void setCompletionTokens(int v) { this.completionTokens = v; }
public int getTotalTokens() { return totalTokens; }
public void setTotalTokens(int v) { this.totalTokens = v; }
public double getCostEur() { return costEur; }
public void setCostEur(double v) { this.costEur = v; }
}

112
dc-backend/src/main/java/de/frigosped/dc/service/AiCostTrackerService.java Normal file
View File

@@ -0,0 +1,112 @@
package de.frigosped.dc.service;
import de.frigosped.dc.client.OrdsClient;
import de.frigosped.dc.model.AiCostRequest;
import dev.langchain4j.model.output.TokenUsage;
import jakarta.enterprise.context.ApplicationScoped;
import jakarta.inject.Inject;
import org.eclipse.microprofile.config.inject.ConfigProperty;
import org.eclipse.microprofile.rest.client.inject.RestClient;
import org.jboss.logging.Logger;
import java.util.concurrent.CompletableFuture;
/**
* Erfasst Token-Verbrauch und Kosten für jeden KI-Aufruf und persistiert
* sie asynchron via ORDS in dc_ai_cost_log.
*
* Aufruf-Muster (qualifier = "main" | "catalog" | "ocr"):
* costTracker.track("main", "TRANSLATE", projectId, null, response.tokenUsage());
*/
@ApplicationScoped
public class AiCostTrackerService {
private static final Logger LOG = Logger.getLogger(AiCostTrackerService.class);
@ConfigProperty(name = "dc.ai.provider", defaultValue = "ollama")
String provider;
// Modellnamen je Provider
@ConfigProperty(name = "dc.ai.ollama.main.model", defaultValue = "") String ollamaMainModel;
@ConfigProperty(name = "dc.ai.ollama.catalog.model", defaultValue = "") String ollamaCatalogModel;
@ConfigProperty(name = "dc.ai.ollama.ocr.model", defaultValue = "") String ollamaOcrModel;
@ConfigProperty(name = "dc.ai.mistral.main.model", defaultValue = "mistral-small-latest") String mistralMainModel;
@ConfigProperty(name = "dc.ai.mistral.catalog.model", defaultValue = "mistral-small-latest") String mistralCatalogModel;
@ConfigProperty(name = "dc.ai.mistral.ocr.model", defaultValue = "pixtral-12b-2409") String mistralOcrModel;
// Preise in EUR pro 1 Mio. Tokens (Mistral; Ollama = 0)
@ConfigProperty(name = "dc.ai.mistral.main.price-input-per-1m-eur", defaultValue = "0.092") double mistralMainInput;
@ConfigProperty(name = "dc.ai.mistral.main.price-output-per-1m-eur", defaultValue = "0.276") double mistralMainOutput;
@ConfigProperty(name = "dc.ai.mistral.catalog.price-input-per-1m-eur", defaultValue = "0.092") double mistralCatalogInput;
@ConfigProperty(name = "dc.ai.mistral.catalog.price-output-per-1m-eur",defaultValue = "0.276") double mistralCatalogOutput;
@ConfigProperty(name = "dc.ai.mistral.ocr.price-input-per-1m-eur", defaultValue = "0.138") double mistralOcrInput;
@ConfigProperty(name = "dc.ai.mistral.ocr.price-output-per-1m-eur", defaultValue = "0.138") double mistralOcrOutput;
@Inject
@RestClient
OrdsClient ordsClient;
/**
* @param qualifier "main" | "catalog" | "ocr"
* @param operation "TRANSLATE" | "EVALUATE" | "OCR" | "CATALOG"
* @param projectId Projekt-ID (null bei Katalog-Operationen)
* @param catalogId Katalog-ID (null bei Prüf-Operationen)
* @param tokenUsage aus response.tokenUsage() – null wird als 0 behandelt
*/
public void track(String qualifier, String operation,
Long projectId, Long catalogId, TokenUsage tokenUsage) {
int promptTokens = safeCount(tokenUsage != null ? tokenUsage.inputTokenCount() : null);
int completionTokens = safeCount(tokenUsage != null ? tokenUsage.outputTokenCount() : null);
int totalTokens = safeCount(tokenUsage != null ? tokenUsage.totalTokenCount() : null);
double costEur = calcCost(qualifier, promptTokens, completionTokens);
AiCostRequest req = new AiCostRequest();
req.setProvider(provider);
req.setModelName(resolveModel(qualifier));
req.setOperation(operation);
req.setProjectId(projectId);
req.setCatalogId(catalogId);
req.setPromptTokens(promptTokens);
req.setCompletionTokens(completionTokens);
req.setTotalTokens(totalTokens);
req.setCostEur(costEur);
CompletableFuture.runAsync(() -> {
try {
ordsClient.logAiCost(req);
LOG.debugf("Kosten erfasst: %s/%s %.6f EUR (%d tokens)",
qualifier, operation, costEur, totalTokens);
} catch (Exception e) {
LOG.warnf("Kostenverfolgung fehlgeschlagen (%s/%s): %s",
qualifier, operation, e.getMessage());
}
});
}
private String resolveModel(String qualifier) {
boolean isMistral = "mistral".equalsIgnoreCase(provider);
return switch (qualifier) {
case "catalog" -> isMistral ? mistralCatalogModel : ollamaCatalogModel;
case "ocr" -> isMistral ? mistralOcrModel : ollamaOcrModel;
default -> isMistral ? mistralMainModel : ollamaMainModel;
};
}
private double calcCost(String qualifier, int prompt, int completion) {
if (!"mistral".equalsIgnoreCase(provider)) return 0.0;
double inPrice, outPrice;
switch (qualifier) {
case "ocr" -> { inPrice = mistralOcrInput; outPrice = mistralOcrOutput; }
case "catalog" -> { inPrice = mistralCatalogInput; outPrice = mistralCatalogOutput; }
default -> { inPrice = mistralMainInput; outPrice = mistralMainOutput; }
}
return (prompt / 1_000_000.0) * inPrice
+ (completion / 1_000_000.0) * outPrice;
}
private int safeCount(Integer value) {
return value != null ? value : 0;
}
}

9
dc-backend/src/main/java/de/frigosped/dc/service/CatalogGenerationService.java
View File

@@ -56,6 +56,9 @@ public class CatalogGenerationService {
@Named("catalog")
ChatLanguageModel catalogModel;
@Inject
AiCostTrackerService costTracker;
@Inject
ObjectMapper objectMapper;
@@ -106,7 +109,7 @@ public class CatalogGenerationService {
catalogId, chunkNr, totalChunks, chunks.get(i).length());
GeneratedCatalogStructure chunkResult =
generateStructureFromChunk(catalogName, chunks.get(i), chunkNr, totalChunks);
generateStructureFromChunk(catalogName, chunks.get(i), chunkNr, totalChunks, catalogId);
if (chunkResult == null || chunkResult.getCategories() == null) {
LOG.warnf("[Katalog %d] Chunk %d lieferte kein Ergebnis – übersprungen",
@@ -316,7 +319,8 @@ public class CatalogGenerationService {
private GeneratedCatalogStructure generateStructureFromChunk(String catalogName,
String chunkText,
int chunkNr,
int totalChunks) {
int totalChunks,
long catalogId) {
List<ChatMessage> messages = List.of(
SystemMessage.from(buildChunkSystemPrompt(chunkNr, totalChunks)),
UserMessage.from(
@@ -327,6 +331,7 @@ public class CatalogGenerationService {
);
Response<AiMessage> response = catalogModel.generate(messages);
costTracker.track("catalog", "CATALOG", null, catalogId, response.tokenUsage());
String rawJson = response.content().text();
return parseCatalogStructure(rawJson, catalogName + " Chunk " + chunkNr);
}

6
dc-backend/src/main/java/de/frigosped/dc/service/CheckOrchestrationService.java
View File

@@ -151,11 +151,11 @@ public class CheckOrchestrationService {
byte[] fileBytes = fileResp.readEntity(byte[].class);
originalText = documentProcessingService.extractText(
fileBytes, doc.getMimeType(), doc.getFilename());
fileBytes, doc.getMimeType(), doc.getFilename(), projectId);
LOG.debugf("Dokument %d: %d Zeichen extrahiert",
(long) doc.getId(), (long) originalText.length());
translatedText = evaluationService.translate(originalText);
translatedText = evaluationService.translate(originalText, projectId);
LOG.debugf("Dokument %d: Übersetzung fertig (%d Zeichen)",
(long) doc.getId(), (long) translatedText.length());
@@ -204,7 +204,7 @@ public class CheckOrchestrationService {
LOG.debugf("Auswertung: Dok %d × Frage %d",
doc.getId(), question.getQuestionId());
EvaluationResult result = evaluationService.evaluate(question, originalText);
EvaluationResult result = evaluationService.evaluate(question, originalText, projectId);
int deviation = 0;
int warning = 0;

24
dc-backend/src/main/java/de/frigosped/dc/service/DocumentProcessingService.java
View File

@@ -46,6 +46,9 @@ public class DocumentProcessingService {
@Named("ocr")
ChatLanguageModel ocrModel;
@Inject
AiCostTrackerService costTracker;
// =========================================================================
// Öffentliche API
// =========================================================================
@@ -59,13 +62,17 @@ public class DocumentProcessingService {
* @return Markdown-String
*/
public String extractText(byte[] fileBytes, String mimeType, String filename) {
return extractText(fileBytes, mimeType, filename, null);
}
public String extractText(byte[] fileBytes, String mimeType, String filename, Long projectId) {
String effectiveMime = resolveMimeType(mimeType, filename);
LOG.debugf("Extrahiere Text: %s (%s)", filename, effectiveMime);
try {
return switch (effectiveMime) {
case "application/pdf"
-> extractFromPdf(fileBytes);
-> extractFromPdf(fileBytes, projectId);
case "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
"application/msword"
-> extractFromDocx(fileBytes);
@@ -88,7 +95,7 @@ public class DocumentProcessingService {
// PDF: Seiten rendern + KI-OCR
// =========================================================================
private String extractFromPdf(byte[] fileBytes) throws Exception {
private String extractFromPdf(byte[] fileBytes, Long projectId) throws Exception {
LOG.debug("Starte PDF-OCR...");
StringBuilder result = new StringBuilder();
@@ -101,12 +108,11 @@ public class DocumentProcessingService {
LOG.debugf("OCR Seite %d/%d", i + 1, pageCount);
BufferedImage image = renderer.renderImageWithDPI(i, OCR_DPI, ImageType.RGB);
// Bild in PNG-Base64 umwandeln
ByteArrayOutputStream baos = new ByteArrayOutputStream();
ImageIO.write(image, "PNG", baos);
String base64Image = Base64.getEncoder().encodeToString(baos.toByteArray());
String pageText = ocrPage(base64Image, i + 1, pageCount);
String pageText = ocrPage(base64Image, i + 1, pageCount, projectId);
result.append(pageText).append("\n\n");
}
}
@@ -114,10 +120,7 @@ public class DocumentProcessingService {
return result.toString().trim();
}
/**
* Sendet eine PDF-Seite als Base64-Bild an das Vision-Modell.
*/
private String ocrPage(String base64Image, int pageNum, int totalPages) {
private String ocrPage(String base64Image, int pageNum, int totalPages, Long projectId) {
UserMessage message = UserMessage.from(
ImageContent.from(base64Image, "image/png"),
TextContent.from(
@@ -134,7 +137,10 @@ public class DocumentProcessingService {
)
);
return ocrModel.generate(List.of(message)).content().text();
dev.langchain4j.model.output.Response<dev.langchain4j.data.message.AiMessage> response =
ocrModel.generate(List.of(message));
costTracker.track("ocr", "OCR", projectId, null, response.tokenUsage());
return response.content().text();
}
// =========================================================================

10
dc-backend/src/main/java/de/frigosped/dc/service/EvaluationService.java
View File

@@ -39,6 +39,9 @@ public class EvaluationService {
@Named("main")
ChatLanguageModel mainModel;
@Inject
AiCostTrackerService costTracker;
@Inject
ObjectMapper objectMapper;
@@ -51,9 +54,10 @@ public class EvaluationService {
* Wenn der Text bereits auf Deutsch ist, wird er unverändert zurückgegeben.
*
* @param originalText Markdown-Text in Originalsprache
* @param projectId Für Kostenzuordnung
* @return Markdown-Text auf Deutsch
*/
public String translate(String originalText) {
public String translate(String originalText, long projectId) {
if (originalText == null || originalText.isBlank()) return "";
LOG.debugf("Starte Übersetzung (%d Zeichen)...", originalText.length());
@@ -71,6 +75,7 @@ public class EvaluationService {
);
Response<AiMessage> response = mainModel.generate(messages);
costTracker.track("main", "TRANSLATE", projectId, null, response.tokenUsage());
return response.content().text().trim();
}
@@ -88,7 +93,7 @@ public class EvaluationService {
* @param documentText Originaltext des Dokuments (Markdown)
* @return Strukturiertes Auswertungsergebnis
*/
public EvaluationResult evaluate(OrdsQuestion question, String documentText) {
public EvaluationResult evaluate(OrdsQuestion question, String documentText, long projectId) {
LOG.debugf("Werte Frage %d aus: %s",
question.getQuestionId(),
abbreviate(question.getQuestionText(), 80));
@@ -102,6 +107,7 @@ public class EvaluationService {
);
Response<AiMessage> response = mainModel.generate(messages);
costTracker.track("main", "EVALUATE", projectId, null, response.tokenUsage());
String rawText = response.content().text();
return parseEvaluationResult(rawText, question);

44
dc-backend/src/main/resources/application.properties
View File

@@ -12,25 +12,55 @@ quarkus.http.port=8090
dc.api.key=${DC_API_KEY:}
# =============================================================================
# KI-Modelle (Ollama-kompatibler Endpunkt)
# KI-Provider: ollama | mistral
# =============================================================================
dc.ai.provider=mistral
# =============================================================================
# Ollama-Konfiguration (aktiv wenn dc.ai.provider=ollama)
# =============================================================================
dc.ai.ollama.base-url=https://ollama.aquantico.de
dc.ai.ollama.api-key=324GF44-50AA-4B57-9386-K435DLJ764DFR
dc.ai.ollama.main.model=qwen3.6:35b-a3b-q4_K_M
dc.ai.ollama.catalog.model=qwen3.6:35b-a3b-q4_K_M
dc.ai.ollama.ocr.model=qwen3.6:35b-a3b-q4_K_M
# =============================================================================
# Mistral-Konfiguration (aktiv wenn dc.ai.provider=mistral)
# =============================================================================
dc.ai.mistral.api-key=${MISTRAL_API_KEY:2AkoohJC8MjfjhfLGObQESZs9jxbDnc0}
dc.ai.mistral.main.model=mistral-small-latest
dc.ai.mistral.catalog.model=mistral-small-latest
# pixtral-12b-2409 unterstützt Vision (PDF-Seiten als Bild)
dc.ai.mistral.ocr.model=pixtral-12b-2409
# Preise in EUR pro 1 Mio. Tokens (Stand 2025, 1 USD = 0.92 EUR)
# mistral-small-latest: $0.10/$0.30 pro 1M tokens
dc.ai.mistral.main.price-input-per-1m-eur=0.092
dc.ai.mistral.main.price-output-per-1m-eur=0.276
dc.ai.mistral.catalog.price-input-per-1m-eur=0.092
dc.ai.mistral.catalog.price-output-per-1m-eur=0.276
# pixtral-12b-2409: $0.15/$0.15 pro 1M tokens
dc.ai.mistral.ocr.price-input-per-1m-eur=0.138
dc.ai.mistral.ocr.price-output-per-1m-eur=0.138
# =============================================================================
# Gemeinsame Modell-Parameter (für beide Provider)
# =============================================================================
dc.ai.base-url=https://ollama.aquantico.de
dc.ai.api-key=324GF44-50AA-4B57-9386-K435DLJ764DFR
# Hauptmodell: Auswertung + Übersetzung
dc.ai.main.model=qwen3.6:35b-a3b-q4_K_M
dc.ai.main.timeout=300s
dc.ai.main.max-retries=2
dc.ai.main.temperature=0.1
# Katalog-Generierungsmodell: längerer Timeout für vollständige Katalog-Analyse
dc.ai.catalog.model=qwen3.6:35b-a3b-q4_K_M
# Katalog-Generierungsmodell
dc.ai.catalog.timeout=1200s
dc.ai.catalog.max-retries=1
dc.ai.catalog.temperature=0.1
# OCR-Modell: Bildtext-Extraktion aus PDF-Seiten
dc.ai.ocr.model=qwen3.6:35b-a3b-q4_K_M
dc.ai.ocr.timeout=120s
dc.ai.ocr.max-retries=2

44
dc-backend/target/classes/application.properties
View File

@@ -12,25 +12,55 @@ quarkus.http.port=8090
dc.api.key=${DC_API_KEY:}
# =============================================================================
# KI-Modelle (Ollama-kompatibler Endpunkt)
# KI-Provider: ollama | mistral
# =============================================================================
dc.ai.provider=mistral
# =============================================================================
# Ollama-Konfiguration (aktiv wenn dc.ai.provider=ollama)
# =============================================================================
dc.ai.ollama.base-url=https://ollama.aquantico.de
dc.ai.ollama.api-key=324GF44-50AA-4B57-9386-K435DLJ764DFR
dc.ai.ollama.main.model=qwen3.6:35b-a3b-q4_K_M
dc.ai.ollama.catalog.model=qwen3.6:35b-a3b-q4_K_M
dc.ai.ollama.ocr.model=qwen3.6:35b-a3b-q4_K_M
# =============================================================================
# Mistral-Konfiguration (aktiv wenn dc.ai.provider=mistral)
# =============================================================================
dc.ai.mistral.api-key=${MISTRAL_API_KEY:2AkoohJC8MjfjhfLGObQESZs9jxbDnc0}
dc.ai.mistral.main.model=mistral-small-latest
dc.ai.mistral.catalog.model=mistral-small-latest
# pixtral-12b-2409 unterstützt Vision (PDF-Seiten als Bild)
dc.ai.mistral.ocr.model=pixtral-12b-2409
# Preise in EUR pro 1 Mio. Tokens (Stand 2025, 1 USD = 0.92 EUR)
# mistral-small-latest: $0.10/$0.30 pro 1M tokens
dc.ai.mistral.main.price-input-per-1m-eur=0.092
dc.ai.mistral.main.price-output-per-1m-eur=0.276
dc.ai.mistral.catalog.price-input-per-1m-eur=0.092
dc.ai.mistral.catalog.price-output-per-1m-eur=0.276
# pixtral-12b-2409: $0.15/$0.15 pro 1M tokens
dc.ai.mistral.ocr.price-input-per-1m-eur=0.138
dc.ai.mistral.ocr.price-output-per-1m-eur=0.138
# =============================================================================
# Gemeinsame Modell-Parameter (für beide Provider)
# =============================================================================
dc.ai.base-url=https://ollama.aquantico.de
dc.ai.api-key=324GF44-50AA-4B57-9386-K435DLJ764DFR
# Hauptmodell: Auswertung + Übersetzung
dc.ai.main.model=qwen3.6:35b-a3b-q4_K_M
dc.ai.main.timeout=300s
dc.ai.main.max-retries=2
dc.ai.main.temperature=0.1
# Katalog-Generierungsmodell: längerer Timeout für vollständige Katalog-Analyse
dc.ai.catalog.model=qwen3.6:35b-a3b-q4_K_M
# Katalog-Generierungsmodell
dc.ai.catalog.timeout=1200s
dc.ai.catalog.max-retries=1
dc.ai.catalog.temperature=0.1
# OCR-Modell: Bildtext-Extraktion aus PDF-Seiten
dc.ai.ocr.model=qwen3.6:35b-a3b-q4_K_M
dc.ai.ocr.timeout=120s
dc.ai.ocr.max-retries=2

BIN
dc-backend/target/classes/de/frigosped/dc/ai/AiModelProducer.class
View File

Binary file not shown.

BIN
dc-backend/target/classes/de/frigosped/dc/client/OrdsClient.class
View File

Binary file not shown.

BIN
dc-backend/target/classes/de/frigosped/dc/service/CheckOrchestrationService.class
View File

Binary file not shown.

BIN
dc-backend/target/classes/de/frigosped/dc/service/DocumentProcessingService.class
View File

Binary file not shown.

BIN
dc-backend/target/classes/de/frigosped/dc/service/EvaluationService.class
View File

Binary file not shown.

2
dc-backend/target/maven-status/maven-compiler-plugin/compile/default-compile/createdFiles.lst
View File

@@ -9,6 +9,7 @@ de/frigosped/dc/service/MarkdownExportService.class
de/frigosped/dc/service/DocumentProcessingService.class
de/frigosped/dc/model/CatalogGenerateResponse.class
de/frigosped/dc/service/MarkdownExportService$Seg.class
de/frigosped/dc/model/AiCostRequest.class
de/frigosped/dc/model/OrdsCatalogStatusResponse.class
de/frigosped/dc/model/OrdsDocumentTexts.class
de/frigosped/dc/resource/CatalogResource.class
@@ -21,3 +22,4 @@ de/frigosped/dc/client/OrdsLoggingFilter.class
de/frigosped/dc/service/MarkdownExportService$PdfWriter.class
de/frigosped/dc/security/ApiKeyFilter.class
de/frigosped/dc/model/GeneratedCatalogStructure$GeneratedQuestion.class
de/frigosped/dc/service/AiCostTrackerService.class

2
dc-backend/target/maven-status/maven-compiler-plugin/compile/default-compile/inputFiles.lst
View File

@@ -1,6 +1,7 @@
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/ai/AiModelProducer.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/client/OrdsClient.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/client/OrdsLoggingFilter.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/model/AiCostRequest.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/model/CatalogGenerateResponse.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/model/EvaluationResult.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/model/ExportRequest.java
@@ -23,6 +24,7 @@
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/resource/CheckResource.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/resource/ExportResource.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/security/ApiKeyFilter.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/service/AiCostTrackerService.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/service/CatalogGenerationService.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/service/CheckOrchestrationService.java
/mnt/c/source/Frigosped/Dokumenten-Check/dc-backend/src/main/java/de/frigosped/dc/service/DocumentProcessingService.java